Applying N at optimal rates has the potential to improve N use efficiency, crop yield, and profitability as well as to reduce environmental impacts (Kyveryga et al., 2009; Wang et al., 2003); however, guidelines on adjusting optimal N rates based on soil and weather conditions are lacking (Tremblay, 2004). Many current N management decisions disregard the effect of interannual temperature and rainfall variations on soil N mineralization (Raun et al., 2005; Melkonian et al., 2007; Shanahan et al., 2008). Weather is a major determinant of soil biological activity, including the decomposition of soil organic matter, and climatic conditions can vary significantly in space and time across North American regions (Bolinder et al., 2007; van Es et al., 2007; Lokupitiya et al., 2010).
Crop growth models can be used to assess optimal N rates; however, the predictions are fairly imprecise and vary substantially among these models (Kyveryga et al., 2007; Naud et al., 2008). Under site-specific N fertilization strategies, some researchers have recommended applying more N to historically high-yielding areas and less to low-yielding areas, whereas others advocate the opposite approach (James and Godwin, 2003). Producers typically apply rates of N fertilizer that they consider sufficient to support near-maximum yields.
The influence of soil texture on N response is well documented, but contradictory results exist as well. In wet climates, yield is generally higher (and N response lower) on coarse-textured soils than on fine-textured soils (Tremblay et al., 2011). In arid climates, higher crop yields are often obtained in clayey soils (higher water-holding capacity) than in sandy soils (Armstrong et al., 2009). Approaches based solely on yield maps do not provide robust information for the determination of management zones (Kitchen et al., 2008). Topography, remote sensing, and soil apparent electrical conductivity have also been used with varying degrees of success to delineate zones of differential response to N rates (Cambouris et al., 2008; Shanahan et al., 2008; Tremblay et al., 2011); however, these methodologies also disregard the effects of weather in determining crop N fertilizer requirements.
Soil properties (including texture, water-holding capacity, and fertility) strongly affect soil N availability and crop yield (Zhu et al., 2009; Armstrong et al., 2009). Some studies have reported that corn N response is only marginally affected by soil texture and that yearly variation has a more pronounced effect than soil spatial variability (van Es et al., 2005; Tremblay and Bélec, 2006; Kyveryga et al., 2009). Precipitation and thermal units have been found to significantly affect soil mineral N and thus corn response to N (Tremblay, 2004; Tremblay and Bélec, 2006; Shanahan et al., 2008; Kyveryga et al., 2007). Shahandeh et al. (2011) showed that corn grain yield was either negatively or positively related to clay content depending on precipitation. Anwar et al. (2009) reported that crop growth is highly sensitive to factors that vary in both space (soil properties) and time (rainfall and temperature). Interactions among these factors control water and nutrient availability as well as N mineralization during the growing season (Schröder et al., 2000; Kay et al., 2006). It follows that proper N management should consider soil texture as well as seasonal conditions of temperature and precipitation (Derby et al., 2005; Shanahan et al., 2008; Sogbedji et al., 2001). Based on models for corn crop growth and N uptake, soil N transformations, and water and N transport, Melkonian et al. (2007) have developed the Precision Nitrogen Management model to improve N use efficiency and reduce N losses. This model uses soil textural class, soil organic matter, weather data, and other information about management practices such as tillage, plant density, and rotations to determine in-season N recommendations for the northeastern United States.
Before the 1990s, data from multiple studies were combined in a narrative review in which a researcher would summarize the response curves of individual studies to reach a conclusion. This approach assigns the same weight to each study and captures the solution as the number of studies increases (Borenstein et al., 2009). Meta-analysis is a statistical method that synthesizes the results of a set of studies. It is used in many fields of research such as medicine, social science, and ecology. Meta-analysis is commonly used to assess the consistency of treatment effect (also called effect size) across a series of studies or experiments. If the treatment effect varies from one study to the next (which is often the case for N fertilization studies), meta-analysis can be applied to assess the levels of effect for subgroups and thus identify factors associated with the magnitude of the effect sizes (Borenstein et al., 2009). Meta-analysis is a systematic method for combining the results from a series of studies and addressing apparently conflicting findings by identifying potential explanatory variables (Olkin and Shaw, 1995). Meta-analysis is suitable for agronomic research in which several investigators have examined similar problems and generated substantial information, sometimes characterized by heterogeneity and contradictions. Valkama et al. (2009) studied the response to P fertilizer application rates in 400 experiments conducted during an 80-yr period in Finland and used plant groups, soil properties, and cultivation zones to explain the differences. Tonitto et al. (2006) conducted a meta-analysis of experiments reported in the literature to compare crop yield response to N fertilization and soil N status as affected by climate, soil texture, and management practices. Chivenge et al. (2011) conducted a meta-analysis of 57 studies concerning smallholder farms in sub-Saharan Africa and found that corn response to added N is higher in clay soils than loam and sand and also higher for higher annual precipitation. Xia and Wan (2008) studied the response of 456 plant species to N additions in their meta-analysis of a log-ratio of plant biomass and tissue from 304 published studies. They used a mixed (random) model and a subgroup heterogeneity analysis and found that N response increased with temperature and annual precipitation.
There is a need to learn more about the effect of soil properties and weather conditions on soil N dynamics and crop response to N to develop algorithms that can be used to recommend appropriate in-season N application rates (Khosla et al., 2002; Chang et al., 2003; Franzen, 2004). With a better understanding of the spatial and temporal variability of N levels in soil and plant N uptake, N management practices could be adjusted to ensure that both economic and environmental objectives are met (Jemison and Fox, 1994; Shahandeh et al., 2011; Shanahan et al., 2008). The high spatial and temporal variability in yield response to N fertilizer that is observed in individual yield response trials leads to a high degree of uncertainty when estimating economically optimum rates of N for a group of trials and when extrapolating these rates from one location to another (Kyveryga et al., 2009). So far, no studies have quantified the effect on N response of combined soil and weather conditions across a number of years in a large geographic area devoted to corn production in North America. Furthermore, it is difficult to have a uniform data set that considers identical treatments for a given region.
The aim of this study was to quantify the effects of soil characteristics and weather properties and the interactions among these factors on corn response to N applications. A meta-analysis was conducted using mirror studies undertaken in several North American locations between 2006 and 2009 with the same N treatments to address the following questions: (i) to what extent do soil and weather properties affect the corn response to N fertilization; and (ii) how significant are the relationships between corn response and N fertilization in homogeneous classes of soil and weather properties?
Materials and Methods
Site Locations and Soil Properties
Experiments were conducted between 2006 and 2009 on experimental farms in the United States, Mexico, and Canada (Fig. 1) to cover a wide range of soil and climatic conditions. Each site is described in Table 1.
|Study no.||Study||Surface soil texture||Growth stage at sidedressing|
|1||Missouri, Clarkton, 2006 (irrigated)||loamy fine sand||V6|
|2||Oklahoma, Stillwater, 2006||fine sandy loam||V8|
|3||Oklahoma, Stillwater, 2008||fine sandy loam||V8|
|4||Illinois, Dixon Springs 2006||silt loam||V5|
|5||Illinois, Dixon Springs, 2007||silt loam||V8|
|6||Kansas, Manhattan, 2006||silt loam||V9|
|7||Kansas, Manhattan, 2008||silt loam||V9|
|8||Missouri, Centralia, 2006 (irrigated)||silt loam||V10|
|9||Missouri, Miami, 2006||silt loam||V9|
|10||Missouri, Portageville, 2006 (irrigated)||silt loam||V6|
|11||Missouri, Centralia, 2007||silt loam||V9|
|12||Missouri, Portageville, 2008 (irrigated)||silt loam||V8|
|13||Nebraska, Shelton, 2006 (irrigated)||silt loam||V8|
|14||Ohio, Wooster 1, 2006||silt loam||V4|
|15||Ohio, Wooster 2, 2006||silt loam||V8|
|16||Ohio, Wooster 1, 2007||silt loam||V4|
|17||Ohio, Wooster 2, 2007||silt loam||V8|
|18||Ohio, Wooster 1, 2008||silt loam||V4|
|19||Ohio, Wooster 2, 2008||silt loam||V8|
|20||Ohio, Wooster 1, 2009||silt loam||V4|
|21||Ohio, Wooster 2, 2009||silt loam||V8|
|22||Ohio, Western 1, 2006||sandy loam/sandy clay loam||V4|
|23||Ohio, Western 2, 2006||sandy loam/sandy clay loam||V8|
|24||Ohio, Western 1, 2007||sandy loam/sandy clay loam||V4|
|25||Ohio, Western 2, 2007||sandy loam/sandy clay loam||V8|
|26||Ohio, Western 1, 2008||sandy loam/sandy clay loam||V4|
|27||Ohio, Western 2, 2008||sandy loam/sandy clay loam||V8|
|28||Ohio, Western 1, 2009||sandy loam/sandy clay loam||V4|
|29||Ohio, Western 2, 2009||sandy loam/sandy clay loam||V8|
|30||Virginia, Blacksburg, 2006||loam||V6|
|31||Virginia, ATD, 2007||loam||V6|
|32||Virginia, BHD, 2007||loam||V6|
|33||Virginia, MCD, 2007||loam||V6|
|34||Virginia, Varina, 2007||loam||V4|
|35||Quebec, L’Acadie, 2007||loam||V8|
|36||Quebec, L’Acadie 2007 (irrigated)||loam||V8|
|37||Quebec, L’Acadie, 2006||clay loam||V6|
|38||Quebec, L’Acadie, 2008||clay loam||V7|
|39||Quebec, L’Acadie, 2009||clay loam||V6|
|40||Ohio, Northwest 1, 2006||silty clay loam||V8|
|41||Ohio, Northwest 1, 2007||silty clay loam||V4|
|42||Ohio, Northwest 2, 2007||silty clay loam||V8|
|43||Ohio, Northwest 1, 2009||silty clay loam||V4|
|44||Ohio, Northwest 2, 2009||silty clay loam||V8|
|45||Missouri, Wilton, 2006||silty clay||V10|
|46||Missouri, Rocheport, 2007||silty clay||V9|
|47||Mexico, Cd Obregón, 2007||clay||V7|
|48||Mexico, MC, 2007||clay||V7|
|49||Mexico, MP, 2007||clay||V7|
|50||Mexico, MC, 2008||clay||V7|
|51||Mexico, MP, 2008||clay||V7|
Soil textures were first grouped into three categories, in keeping with the approach used by Tonitto et al. (2006): fine textures (clay, silty clay, silty clay loam, and clay loam), medium textures (loam and silt loam), and coarse textures (sandy loam, sandy clay loam, loamy fine sand, and fine sandy loam). Because medium and coarse textures showed a similar N response behavior (data not shown), however, only two classes were retained: (i) fine-textured soils, including clay, silty clay, silty clay loam, and clay loam textures; and (ii) medium to coarse textures (hereafter called medium for greater simplicity), including loam, silt loam, sandy loam, sandy clay loam, loamy fine sand, and fine sandy loam textures. Fifteen of the 51 studies involved fine-textured soils and 36 involved medium-textured soils (Table 1). In this classification, the soil was considered fine textured above a clay content threshold of 30%.
Nitrogen Treatments and Replications
An important characteristic of this research was that the same N rates were applied in all 51 mirror studies. Nine N rate treatments were randomized within three or four blocks in each field. The control treatment received 0 kg N ha−1. Seven other treatments consisted of the same amount of N as a starter (36 kg ha−1) at sowing and increasing N rates at sidedressing (in-season N rates, ISNR): 0, 27, 54, 80, 107, 134, and 161 kg ha−1 applied according to local timing practices at growing stages ranging from V4 to V10 (median: V7) with incidentally 10 pairs of studies with growth stages V4 and V8 at sidedressing in Ohio in the same years at the same sites (Table 1). The last N treatment consisted of 178 kg ha−1 applied at sowing, with no N fertilizer at sidedressing; it is referred to as the N-rich rate. This treatment provided the opportunity to examine the effect of weather on a high N rate applied early in the season.
Weather Data and Weather Parameters
Daily rainfall (Rain) data and daily minimum and maximum temperatures (Tmin and Tmax) were collected for each site-year. For practical reasons, these simple and easily available data were selected to calculate corn heat units (CHU; Bootsma et al., 2005), cumulative precipitation (PPT), and the Shannon diversity index (SDI; Bronikowski and Webb, 1996). The SDI was used to assess the distribution of rainfall during a given period. These weather parameters were calculated using the equations presented in Table 2. Cumulative CHU values were computed using daily maximum and minimum temperatures; PPT and SDI were calculated from the daily rainfall data (Table 2).
|Corn heat units (CHU)||CHU = Σ(Ymax + Ymin)/2, where Ymax and Ymin are the contributions to CHU from daily maximum (Tmax, up to 30°C) and minimum (Tmin) air temperatures, respectively: Ymax = 3.33(Tmax – 10.0) – 0.084(Tmax – 10.0)2, if Tmax < 10.0, Ymax = 0.0; Ymin = 1.8(Tmin – 4.44), if Tmin < 4.44, Ymin = 0.0|
|Cumulative precipitation (PPT)||PPT = Σ(Rain), where Rain is the daily rainfall (mm)|
|Precipitation evenness: Shannon diversity index (SDI)||SDI = [–Σpi ln(pi)]/ln(n), where pi = Rain/PPT is the fraction of daily rainfall relative to the total rainfall in a given time period and n is the number of days in that period. SDI = 1 implies complete evenness (i.e., equal amounts of rainfall in each day of the period); SDI = 0 implies complete unevenness (i.e., all rain in 1 d)|
We also proposed a parameter representing optimal water availability (abundant rainfall, well distributed in time). We defined abundant and well-distributed rainfall (AWDR) asFour examples of weather data and derived parameters are given in Fig. 2. Water provided as irrigation (NO3–N content not assessed) was considered equivalent to natural rainfall. This assumption was validated by conducting a meta-analysis on the responses to N rate of irrigated and unirrigated sites under the same soil and rainfall conditions, which revealed no significant difference between the presence and absence of irrigation (data not shown).
The time period covered by weather parameters overlapped the date of N sidedressing (SD). To determine the period during which a weather parameter is most closely related to the N response (or response ratio, RR = YieldNrate/Yieldcontrol), the weather parameters were tested for periods from nb days before SD to na days after SD (with nb and na varying between 1 and 35 d). The optimal period for any weather parameter was the one that maximized the difference in N response across N rates. Thus, for CHU, PPT, SDI, and AWDR, we had to find [nb, na] pairs that maximized the contrast between the two classes of global effect size () across studies and N rates. The global effect size is defined aswhere K is the number of studies and is calculated for the high and low classes of each weather parameter (CHU, PPT, SDI, and AWDR) and each [nb, na] pair. These low and high classes were determined by histogram-based thresholding using the Otsu method (Otsu, 1979), which consists in maximizing the between-class variance (and minimizing the within-class variance) to get the optimum threshold separating the two classes.
In contrast with low CHU, high CHU before the SD period led to a higher (Fig. 3a). The period ranging from 30 d before SD (nb = 30) to 15 d after SD (na = 15) was therefore selected. Rainfall properties (PPT, SDI, and their product, AWDR) were more crucial for long periods after SD, while the period before SD was less important (Fig. 3b, 3c, and 3d). For rainfall properties, a critical period from nb = 15 d to na = 30 d was selected. The testing of alternative periods such as SD – 30 to SD, SD to SD + 30, SD – 30 to SD + 30, SD – 20 to SD, SD to SD + 20, SD – 20 to SD + 20, SD – 10 to SD, SD to SD + 10, and SD – 10 to SD + 10 resulted in either not significant or less significant differences between low and high classes (for CHU, PPT, SDI, and AWDR) than the ones obtained with the periods selected (nb = 30 d to na = 15 d for CHU and nb = 15 d to na = 30 d for PPT, SDI, and AWDR).
The CHU, PPT, SDI, and AWDR classes were separated into low and high subclasses for the periods of maximum effect on N response for each weather parameter using the Otsu histogram thresholding method. The thresholds between the low and high subclasses are as follows: 1160 for CHU (nb = 30 d, na = 15 d); 180 mm for PPT (nb = 15 d, na = 30 d); 0.55 for SDI (nb = 15 d, na = 30 d); and 99 for AWDR (nb = 15 d, na = 30 d). Low AWDR could be considered as suboptimal rainfall (rare and sparse) and high AWDR as optimal rainfall (abundant and well distributed).
The distribution of studies in the [CHU, AWDR] space (Fig. 4) shows that several studies with both fine- and medium-textured soils can be found for all combinations of CHU–AWDR classes except the high AWDR–high CHU subgroup, for which only one study was conducted on a fine-textured soil.
The meta-analysis performed in this research was based on the principles described in detail by Borenstein et al. (2009) and summarized below.
The effect size, Y, is a value that reflects the magnitude of the treatment effect. The outcome (in our case, corn grain yield at 14.5% moisture, in Mg ha−1) is measured on a physical scale, and the effect size is expressed as a RR, which is the ratio of the yield obtained for various N rates (YieldNrate) to the yield measured for the N rate = 0 plots (Yieldcontrol). Thus, for each study i and each N rate r:The overbars in Eq.  indicate that the yields are averaged across the replicates. The logarithmic scale is used to maintain symmetry (Tonitto et al., 2006) and allow the addition of effect sizes.
The replicates were also used to assign a weight to the trials in each study and to each N rate. This weight was assumed to be inversely proportional to the variance Vyi,r (within-study variance) of the yields measured in replicates of any study i at any N rate r. Because two treatments are involved in the definition of the effect size (N rate treatment and control), the variance of the effect size is the pooled (combined) variance of these two groups:whereand nNrate and ncontrol are the number of replicates (sample size) of the two groups.
The weight, Wi,r, assigned to each study i for each N rate r is inversely proportional to the variance:This summary effect across studies is the weighted mean of effects:where K is the number of studies. The weighted mean, µr, is calculated for each N rate r.
The analysis of the effect size requires a mathematical model. While the N treatment effect Y can vary from one study to another depending (among other things) on N rate, soil properties, and weather conditions, a variable-effect (also called random-effect) model is used to consider both within-study variance and between-studies variance. We considered the observed effect size Yi,r for a given study i at a given N rate r, which varies from the overall mean µr by a deviation ξi,r, which reflects the variability of the effect size across the studies, and a sampling error εi,r:Thus, for each N rate r, the observed effect size Yi,r varies from its true value θi,r = µr +ξi,r with an error εi,r. The analysis of the heterogeneity of the studies (the magnitude of ξi,r) allows us to identify subgroups characterized by the same treatment effects. This analysis is performed by estimating the two components of the observed variance (Q): the between-studies variance [T2 = var(θ )] and the within-study variance [var(ε)]. For each N rate r, the observed variance is calculated by assigning a weight, Wi,r, to each study i:Because Wi,r is the inverse of the variance of Yi,r, Qr is a standardized measure not affected by the metric of Yi,r. To partition Qr, we assumed that, at a given N rate r, if studies share the same effect size (ξi,r = 0) and all variation is due to sampling errors εi,r, then within studies, the expected value of Qr is equal to the degrees of freedom, df = K – 1 [i.e., var(ε) = df], where K is the number of studies (central limit principle). The excess variation, Qr – df, reflects the differences in the true effects from study to study. Borenstein et al. (2009) proposed two different statistics that can be used to perform a heterogeneity test that is independent of the number of studies (df): T2, the estimated variance of the true effect size, is given bywhereand Ir2, the proportion of the between-studies variance relative to the total variance, is given byThe T statistic is expressed in the same metric as the effect size Y, while I2 is a ratio independent of the metric and the number of studies.
With the variable-effect (random-effect) model, the between-studies variance should be considered in calculating the weights, Wi,r, assuming that the total variance of a study is the sum of the within-study variance, Vyi,r, and the between-studies variance, Tr2:The variable Wi,r is replaced by Wi,r* (Eq. , , and ). The use of the weighted Wi,r* avoids allocating an excessive weight to any study i if its variance Vyi,r, is too small because Tr2 is considered in the definition of Wi,r*.
Meta-analysis is useful for quantifying the extent of the heterogeneity and understanding the underlying causes. The method used to determine if the studies are heterogeneous is based on I2 (proportion of between-studies variance) levels. For each N rate, r, if Ir2 is close to zero (or negative), the groups are considered homogenous: the observed variance is random and due to sampling error. On the other hand, if Ir2 is high, the causes of the variations should be investigated by performing analyses on subgroups using potential explanatory factors. The values 0.25, 0.50, and 0.75 correspond to low, medium, and high I2 levels, respectively (Parent, personal communication, 2012). From this point, the index r will be omitted and I2 will be used for all N rates.
The above-described heterogeneity analysis on all the studies was used to assess:
subgroups of soil textures established from N response behavior across studies;
subgroups of weather conditions established from N response behavior across studies;
subgroups of combined soil textures and weather conditions.
The variance explained by the classification into subgroups is defined as the ratio of explained variance and total variance (Borenstein et al., 2009). Because the explained variance equals total variance minus unexplained variance (within subgroups), the proportion of the variance explained is given bywhere
Raw Yield Data
Figure 5 shows corn grain yields (mean of replicates) in each study by N rate and the corresponding information on soil texture and AWDR–CHU classes. In fine-textured soils, yields never exceeded 12.5 Mg ha−1. Very low yields were found in both medium- and fine-textured soils at low N rates (e.g., Studies 3 and 39) and at high N rates (e.g., Studies 14–15 and 48–49). The N-rich rate (178 kg ha−1, all at sowing) produced a significantly lower yield (according to a t-test) than an equivalent split application (starter 36 + ISNR 134 kg ha−1) in 15% of the cases when AWDR was low (for both fine- and medium-textured soils), 33% of the cases for the medium texture–high AWDR class, and 63% of the cases for the fine texture–high AWDR class. The N-rich rate gave significantly higher yields than starter 36 + ISNR 134 kg ha−1 in only two studies (5 and 33) corresponding to the medium texture–low AWDR condition. The control rate (0 N) produced low yield, particularly under high AWDR and especially in fine-textured soils. Indeed, Yieldcontrol was lower than 5 Mg ha−1 in 9% of the cases for the medium texture–low AWDR class, 27% of the cases for the medium texture–high AWDR class, 14% of the cases for the fine texture–low AWDR class, and 100% of the cases for the fine texture–high AWDR class. In the latter case, the low Yieldcontrol was probably due to N losses, mainly by denitrification, caused by abundant precipitation in poorly drained soils (van Es et al., 2007; Sogbedji et al., 2001).
Figure 5 also shows that both yield and yield response to N were highly variable among studies, but it does not reveal clear relationships between yield, soil texture, and weather. It illustrates the need for a meta-analysis to provide greater weights to more reliable studies with a goal of building homogeneous subgroups with meaningful summary effect sizes.
Meta-analysis of Subgroups
Meta-analysis provided a summary effect size for each subgroup by N rate (N rate = starter + ISNR or N-rich rate). The I2 values indicated some heterogeneity when all studies were grouped together (across N rates, except N rate at ISNR = 0) and justified subgroup analysis (Table 3). Subgroups were formed for soil texture classes, weather parameter classes, and texture–weather class combinations. The I2 values were calculated for each subgroup and each N rate. Negative values of I2 were not set to zero to give a better idea of the relative degree of homogeneity of the subgroups.
|Subgroup†||Nb‡||36 + 0||36 + 27||36 + 54||36 + 80||36 + 107||36 + 134||36 + 161||N rich|
|All studies together|
|All textures, all weather||51||–0.26||0.20||0.28||0.21||0.26||0.27||0.31||0.34|
|Subgroups for texture properties|
|Subgroups for weather properties|
|Subgroups for combined texture and weather properties|
|Fine texture–low CHU||12||0.07||0.43||0.42||0.40||0.29||0.44||0.51||0.56|
|Fine texture–high CHU||3||–0.12||0.16||–0.11||–0.46||0.44||–0.34||–0.16||–0.40|
|Medium texture–low CHU||16||–0.13||–0.01||0.09||–0.13||0.06||–0.06||0.04||0.13|
|Medium texture–high CHU||20||0.39||0.18||0.19||0.14||0.04||0.40||0.07||0.17|
|Fine texture–low AWDR||8||–0.16||0.08||0.26||0.28||0.10||0.25||0.43||0.47|
|Fine texture–high AWDR||7||–0.07||0.09||0.09||0.10||0.11||0.26||0.25||0.16|
|Medium texture–low AWDR||21||–0.66||–0.07||–0.04||–0.10||–0.01||0.14||–0.06||–0.10|
|Medium texture–high AWDR||15||–0.45||0.03||–0.37||–0.91||–0.27||–0.13||–0.09||0.17|
|Subgroups for rainfall and CHU for fine soil textures combined|
|Low AWDR–low CHU||5||–0.22||–0.06||0.26||0.03||0.19||0.11||0.17||0.28|
|Low AWDR–high CHU||2||–0.02||–0.11||0.00||–0.01||0.07||–0.01||0.01||–0.02|
|High AWDR–low CHU||7||–0.15||0.09||0.07||0.14||0.25||0.27||0.30||0.06|
|High AWDR–high CHU||1||–||–||–||–||–||–||–||–|
The tree diagram of the meta-analysis is shown in Fig. 6a and 6b. The subgroups were either combinations of soil texture and CHU (Fig. 6a) or combinations of soil texture and AWDR (Fig. 6b). The AWDR was considered to be more representative of rainfall conditions than PPT or SDI taken alone (see discussion of weather classes below). The effect sizes, Yi, are indicated by circular points with a size (surface) proportional to the weight Wi*. The error bars indicate the standard error (±SEY), which is equal to the root square of Vy*.
The Yi values are characterized by higher dispersion across studies for higher ISNR as well as for the N-rich rate (Fig. 6a and 6b). The I2 for N rate = 36 + 0 kg ha−1 indicated high homogeneity in almost all subgroups (Table 3). This was expected because a higher N rate leads to more variability of the N response depending on the growing conditions in each study (Haberle et al., 2008). This dispersion does not fully explain heterogeneity, however, because it also depends on the variance Vyi (experimental error, indicated by the error bars in the figures). The heterogeneity of the effect sizes described by the between-studies variance, I2, reflected the dispersion of the Yi values (Table 3). The weighted averages of Yi indicated by dotted lines show the different levels of the effect size in each subgroup (Fig. 6a and 6b). The subgroup fine texture–high AWDR shows a higher effect size than the other subgroups across N rates. A more detailed subgroup analysis is presented next.
Soil Texture Classes
Soil texture class (fine or medium) determined the N response to a large extent (Fig. 7). The average (weighted mean) RR was higher in fine texture classes than in medium texture classes, and this difference increased with increasing N rate. The RR also showed greater heterogeneity across studies at higher N rates (Fig. 6a and 6b) as evidenced by higher I2 values (Table 3) and larger error bars (Fig. 7). The heterogeneity test (Table 3) showed that the effect size, Y, was homogeneous in medium-textured soils (I2 ≤ 0.1) but heterogeneous in fine-textured soils (medium to high I2), except with N rate = 0 kg ha−1.
Soils in the medium-texture class tended to show similar responses to added N, with yield gains varying between 40% (RR ? 1.4) and 65% (RR ? 1.65) and marginal improvements above 134 kg N ha−1 (Fig. 7). The fine-texture class was characterized by a much higher RR, reaching 2.7 at the highest N rate. There was too much heterogeneity (Table 3), however, to determine a reliable summary effect size. The variance explained, R2, by soil texture subgrouping did not exceed 10%, mainly because of a large component affecting the fine-texture class. Subgroup analyses should also consider weather parameters.
Corn heat units, PPT, SDI, and the proposed parameter AWDR (= PPT × SDI) can influence N uptake, mineralization, leaching, and denitrification. This study considered the following periods: 30 d before SD to 15 d after SD for CHU; and 15 d before SD to 30 d after SD for PPT, SDI, and AWDR. These periods were chosen because they showed the strongest relationship with effect of in-season added N.
Twenty-eight of the 51 studies had low CHU values and 23 had high CHU values (Table 3). Responses (RR) were slightly higher at high CHU levels than at low CHU levels (Fig. 8a). The difference was small and error bars overlapped, indicating that CHU alone could not explain the effect size variation. Low CHU subgroups were heterogeneous across most N rates (except ISNR = 0). Hence, CHU alone could not capture the response ratios.
The threshold used for cumulative precipitation yielded 28 studies in the low PPT class and 23 in the high PPT class. The high PPT group was characterized by higher response to added N than the low PPT across N rates (Fig. 8b). The difference was proportional to added N. Overall, the difference between high and low PPT classes was larger than in the case of the CHU classes (Fig. 8a). Low PPT conditions were characterized by higher heterogeneity, except in the case of ISNR = 0 (Table 3). The high PPT group was homogenous across most N rates.
The SDI was low in 30 trials and high in 21 trials (Table 3). Low SDI trials were heterogeneous across N rates except at ISNR = 0. The high SDI group was homogeneous for most N rates except ISNR = 107 kg ha−1 and the N-rich rate. Response ratios were higher in the high SDI class than in the low SDI class (Fig. 8c). The difference was greater for higher N rates and was comparable to that for PPT classes. The correlation between SDI and PPT was very low (0.24). This is indicative of the fact that the spread of precipitation with time has an influence of its own on the response to N.
Because high PPT and SDI classes enhanced the RR more than low classes, their product (AWDR) tended to further increase the difference (Fig. 8d). The increase in the RR associated with high AWDR compared with low AWDR was very large and likewise proportional to the N rate. At the two highest N rates, the RR increased from 1.6 at low AWDR to 2.6 at high AWDR. Moreover, the AWDR classes showed lower heterogeneity than PPT or SDI (Table 3). Infrequent rain situations (low SDI) lead to dry soil conditions in which precipitation events, when they occur, are less likely to impact N losses by leaching or denitrification. Frequent rain situations (high SDI) tend to preserve soil moisture, increase the likeliness of leaching or denitrification, and therefore increase the crop response to N fertilization.
In summary, rainfall-based parameters and CHU (to a lesser extent) influenced the RR. The heterogeneity remaining in the effect sizes of the subgroups indicates, however, that neither factor can fully explain the variability in N response. The variance explained, R2, by CHU, PPT, SDI, and AWDR alone did not exceed 12, 8, 4, and 14%, respectively. It was therefore warranted to combine soil texture and weather classes to obtain more homogeneous subgroups.
Combined Soil Texture and Weather Classes
Soil texture and weather classes were combined factorially into texture–weather subgroups. For CHU classes:
fine texture–low CHU
fine texture–high CHU
medium texture–low CHU
medium texture–high CHU
For rainfall classes:
fine texture–low AWDR
fine texture–high AWDR
medium texture–low AWDR
medium texture–high AWDR
Medium-textured soils formed a homogeneous subgroup across N rates (Table 3). Separating the medium-textured class into two CHU classes (medium texture–high CHU [20 studies] and medium texture–low CHU [16 studies]) did not increase the homogeneity. The trials classified into the medium texture–high CHU subgroup generally showed a higher effect size than those in the medium texture–low CHU subgroup (Fig. 9a). This difference was of little significance, however, because at higher N rates, the RR increased from 1.6 in the low CHU subgroup to 1.9 in the high CHU subgroup.
Separating the medium soil texture group into high and low AWDR subgroups improved homogeneity slightly at most N rates (Table 3). The subgroup medium texture–high AWDR showed a higher RR than the medium texture–low AWDR subgroup (Fig. 9b). The difference was greater at higher N rates, where the RR increased from 1.6 at low AWDR to >2 at high AWDR.
The effect size of studies involving fine-textured soils showed a high level of heterogeneity that was reduced by subgroup analysis (Table 3). The fine texture–high CHU subgroup (three studies) was generally homogeneous but the fine texture–low CHU subgroup (12 studies) was not. Figure 9c shows that the RR weighted mean of the fine texture–high CHU studies was lower than that of fine texture–low CHU studies for ISNR > 134 kg ha−1. This difference was not consistent because the subgroups were not homogeneous.
The AWDR reduced the heterogeneity of effect size in the fine-textured soil class, especially for the fine texture–high AWDR subgroup (seven studies), which had a considerably higher RR (reaching 4.5 at high N rates) than the fine texture–low AWDR subgroup (Fig. 9d) and the other texture–weather subgroups (Fig. 9a, 9b, and 9c). The difference increased with increasing N rates. Hence, rainfall patterns had an appreciable effect on the N response in the fine-textured soil class.
The variance explained, R2, was in the interval 10 to 18% for texture–CHU subgrouping and in the interval 25 to 35% for texture–AWDR subgrouping. No relationship between R2 and N rates was observed.
Because heterogeneity was not observed among all weather subgroups for the fine soil texture group, the subgroup analysis was refined by using combined CHU and AWDR classes as follows:
fine texture–low AWDR–low CHU (five studies)
fine texture–low AWDR–high CHU (two studies)
fine texture–high AWDR–low CHU (seven studies)
fine texture–high AWDR–high CHU (one study)
The subgroups were not consistently homogeneous, particularly at high N rates (Table 3). Therefore, greater precision could be attained with this subgrouping of fine soil texture studies. The CHU classes produced different RR levels in both the high and the low AWDR subgroups (Fig. 10). The subgroup fine texture–low AWDR–high CHU gave a higher RR than fine texture–low AWDR–low CHU. For fine texture–high AWDR, however, high CHU gave a lower RR than low CHU for N rates >80 kg ha−1. This was probably due to the very high RR of Studies 38 and 39 (Table 1; Fig. 6a and 6b) in the subgroup fine texture–high AWDR–low CHU compared with Study 37, which alone made up the subgroup fine texture–high AWDR–high CHU. With so few studies in these subgroups, it appeared reasonable to rely on previous findings, which simply indicated that the RR increased with high CHU. In general, in fine-textured soils, it is important to combine CHU and rainfall conditions to better characterize the potential impact of in-season N rates.
The variance explained, R2, by subgroupings involving texture, AWDR, and CHU was in the range 42 to 60%. This level can be considered as very high in comparison with those reported by Kyveryga et al. (2009), who found that the variance of corn yield response to N explained by year did not exceed 25% and that explained by soil properties did not exceed 16%.
Summary of Corn Response in Homogeneous Subgroups
In the medium-textured soil group, both weather parameters (CHU and AWDR) were helpful for forming subgroups of N response. In the case of the fine-texture soil class, it was more effective to use low and high AWDR classes (Fig. 10a).
The medium texture RR was <2.2, and AWDR improved the RR as the N rate increased (Fig. 10b). Splitting the medium texture group into low and high CHU improved the RR to a level close to that observed for the AWDR classes.
In the fine soil texture group, low RR values were obtained when AWDR was low. The RR in the fine texture–low AWDR subgroup showed similar levels to those for the medium texture group. In such circumstances, the N response behavior of the fine soil textures was similar to that of medium textures in the same CHU class.
The fine texture group showed a high RR (3–4.5) when AWDR was high and the N rate ≥116 kg N ha−1 (starter 36 + ISNR 80 kg ha−1) (Fig. 10a). Our data suggest that CHU classes have an inverse effect (high CHU gives a lower RR than low CHU) for the subgroup fine texture–high AWDR, which contained only one study with high CHU. The fine texture–high AWDR subgroup was homogeneous across the CHU classes, even though I2 reached 30% at some N rates (Table 3).
Soil textures (fine or medium) determined the N response to a large extent. Responses to added N were more pronounced for fine texture groups (clay, silty clay, silty clay loam, and clay loam) than for medium texture groups (loam, silt loam, sandy loam, sandy clay loam, and loamy fine sand). It has been reported that corn is more responsive to N fertilization in clayey soils. For instance, Ping et al. (2008) found that corn needed less N fertilizer in sandy soils than in clayey soils. Shahandeh et al. (2011) showed that a higher soil N supply was associated with lower clay content, and lower N supply with higher clay content, probably because of lower N mineralization in clayey soils (Ros et al., 2011; Zhu et al., 2009). We found that corn yields increased by a factor of 1.6 at high N rates in medium-textured soils but by a factor of 2.7 in fine-textured soils.
The CHU parameter had an especially pronounced influence on N rate effects in the period from 30 d before SD to 15 d after SD. The higher relative importance of CHU accumulation before sidedressing than after sidedressing justifies its inclusion in a decision-making system. Rainfall patterns (PPT, SDI, and their product AWDR) had a particularly pronounced influence on size effects in the period from 15 d before SD to 30 d after SD. According to van Es et al. (2007), if high rainfall occurs before SD when the corn plants are still small, it tends to result in N losses and therefore a higher N response. If high rainfall occurs after SD, it mostly results in higher yields (no drought stress) and therefore greater N response as well (Fox and Piekielek, 1998). The greater influence of rainfall patterns following fertilizer N application shows the interest in reliable precipitation forecasts for the prediction of crop N demand. Anwar et al. (2009) expressed the same concern in relation to barley (Hordeum vulgare L.) to predict seasons when the application of N fertilizer would be beneficial. This is less of a problem under irrigated conditions, given that, according to our observations, water provided through irrigation has the same effect on the N response as water received as rainfall.
High CHU tended to enhance the corn response to added N. Higher heat accumulation may lead to more N mineralization from the soil but also to more volatilization, growth, and therefore N uptake from the crop. Current recommendations in Ontario (Ontario Ministry of Agriculture, Food, and Rural Affairs, 2012) suggest a heat unit adjustment because corn in the long-season areas of the province requires more N than in the short-season areas. This may be due to greater moisture stress on the crop in areas with higher average temperatures, which would decrease N use efficiency, or it could be related to differences in soil organic matter content. The adjustment is approximately 1.8 kg N per 100 CHU above or below the base value of 2650. More importantly, higher N rates were more beneficial as PPT increased and was evenly distributed throughout the season. The AWDR was a powerful integrated descriptor of precipitation amount and spread with time. High N rates increased yield by a factor of 2.6 under high AWDR compared with only 1.6 under low AWDR. Ros et al. (2011) explained that mineralizable N is closely related to temperature and moisture content. Xia and Wan (2008) showed in their meta-analysis of 304 published studies that plant responses to N increased with temperature and annual precipitation. According to Tremblay (2004), dry years are characterized by a poor response to N fertilization, and a greater response is observed in wet years. Kyveryga et al. (2009) and Zhu et al. (2009) also found a greater response in years of higher rainfall. Shahandeh et al. (2011) reported that in a wet year, corn response to 180 kg N ha−1 almost doubled in medium-textured soils and tripled in fine-textured soils compared with drier years. This difference was attributed to the decrease in residual soil NO3–N with time under abundant rainfall regimes and to the increase in water available for growth.
In our study, the interactions between soil texture and weather conditions had the greatest influence on the RR. At the lower end of the spectrum were medium-textured soils and the CHU parameter; at the higher end, fine-textured soils under low or high AWDR conditions. Nitrogen applications may increase corn yield in a fine-textured soil by a factor of 1.5 under low AWDR and a factor of 4.5 under high AWDR conditions. In this particular case (fine texture–high AWDR), lower (and not higher) CHU favored the higher response to N rate. Kravchenko et al. (2005) found that the spatial variability of the corn yield response to added N can increase in high-rainfall years. In a meta-analysis of 57 experimental studies in sub-Saharan regions, Chivenge et al. (2011) showed that N response was higher in clay soils than in loam or sand, and also higher at higher annual precipitation levels. According to van Es et al. (2005), the N response was greater in finer textured soils in years with wet springs. Dharmakeerthi et al. (2006) reported that corn N uptake differed at the landscape scale; the magnitude of the difference was greater in seasons with abundant rainfall. The interaction between soil texture and rainfall is probably related to the drainage capacity of the soils (sand has a higher capacity, clay a lower capacity) (Taylor et al., 2003; Shahandeh et al., 2011). Clay retains water for a longer time after precipitation compared with sand (van Es et al., 2005). According to Armstrong et al. (2009), soil water and rainfall affect the relationship between soil texture and the spatial variations in yield through two mechanisms: the first is a complex relationship between subsoil physical-chemical constraints and soil water availability affecting crop growth; the second relates to osmotic effects in the root zone, which increase as the soil water content decreases.
It is noteworthy that the application of all N at sowing tended to be less effective than split applications under high AWDR both in fine-textured soils (Fig. 9d) and in medium-textured soils (Fig. 9b). Thus, as mentioned by van Es et al. (2007), the highest precision in N management may be achieved through in-season N applications that are based on information on late-spring precipitation patterns. This allows the N losses (leaching or denitrification) occurring due to possible excessive rainfall (Kay et al., 2006) to be taken into account. It is worth mentioning that there was generally no influence of growth stage (V4 or V8) on the effectiveness of the application of N fertilizer (Ohio Studies 14 vs. 15, 16 vs. 17, 18 vs. 19, 20 vs. 21, 22 vs. 23, 24 vs. 25, 26 vs. 27, 28 vs. 29, 41 vs. 42, and 43 vs. 44 [Table 1; Fig. 6a and 6b]).
Meta-analysis allowed us to build homogeneous groups based on soil texture, rainfall (AWDR), and CHU classes. Summary effect sizes were computed for each subgroup at each N fertilization rate. The variance explained by this subgrouping reached 42 to 60% (across N rates), which is high considering the large geographic and climatic zones covered by the database. The residual variability within these subgroups is probably not attributable solely to experimental error. Other parameters such as topography, soil organic matter content, previous crop, diseases, insects, NO3 content of the irrigation water, and drainage problems may be involved (Tremblay, 2004; Dharmakeerthi et al., 2006; van Es et al., 2005). The rules derived from this study were based on yield improvement and do not take environmental risks into account. It is generally recognized, however, that N rates resulting in significant yield increases do not lead to unreasonable N losses, particularly when in-season applications are made (Olfs et al., 2005).
In summary, responses to applied N were found to be higher at sites with soils containing >30% clay. Under conditions of high temperatures during the period from 30 d before to 10 d after sidedressing time, the differences should be greater, particularly for fine-textured soils when seasonal rainfall is abundant and well distributed with time (high AWDR). The results may be used for variable N rate management within and between fields and seasons. This study provides guidelines for deriving optimal N rates adapted to local soil texture data and weather conditions (both actual and forecast) both at the regional and field levels. The quantitative information can be easily summarized in an aided decision support system using a set of fuzzy inference system rules from which optimal rates can be calculated, as shown by Tremblay et al. (2010) and Bouroubi et al. (2011).
Several researchers have reported that differential N responses are due to spatial and temporal variations in crop demand and soil N supply and losses; however, N responses have not been quantified according to different soil and weather conditions. This meta-analysis study using a uniform pan-American database provides an approach for deriving in-season N rates that are adapted to soil and weather information. This approach appears particularly well suited to answering questions that cannot easily be addressed using limited experimental data encompassing different soil textures and weather conditions. Soil and weather properties were found to have a fairly pronounced effect on the corn response to N fertilization. Under certain soil and weather conditions (AWDR–CHU subgroups for fine-textured soils), accurate summary effect sizes could not be obtained owing to the limited number of studies. Further studies are necessary to establish reliable patterns for these soil–weather conditions. The measured effects of N rates in relation to soil textures and temperature and precipitation data can be used to derive algorithms permitting in-season N fertilization at levels that are both economical and environmentally benign. If long-term weather forecasts become more reliable, it will be possible to make adjustments not only for past weather conditions but also for those expected up to 30 d after N sidedressing. In the meantime, decisions may be based mainly on historical weather information.