Biophysical and policy factors predict simplified crop rotations in the US Midwest

We compiled a dataset that shows the crop sequence (cash crops only; cover crops are not detected by the Cropland Data Layer) on each field in the study area (see 'datasets' below) and used these sequences as a proxy for crop rotation to derive a novel indicator of rotational complexity that could be applied at the field scale. To date, no metric exists that can supply both the flexibility of quantifying different length rotations that occur in the same time period, and the specificity of operating at the field level.

We adapted the rotational diversity index (d) of previous studies [25, 48] that quantifies the diversity of a known rotation as a function of the length of the rotation (l, in years) and the number of crops in the rotation (n):

$$\begin{equation}d = \sqrt {nl} \end{equation} \tag{ 1 }$$

to accommodate large-scale, remotely-sensed data (see supplement for further details). Instead of trying to identify common rotations of a known length, we observed a fixed-length sequence (in our case, 6 years) of crops in a given field, and replaced rotation length (l) with species turnover from year to year (T₁), and turnover every 2 years (T₂). For example, in a 6 year crop sequence ABABAB, T₁ = 5 and T₂ = 0, while for ABCABC, T₁ = 5 and T₂ = 4. We also added a pseudo-turnover (T_p) equal to the number of times a duplicate perennial appears in a field in two consecutive years (e.g. alfalfa–alfalfa), as perennial crops typically provide soil benefits [49–52] but would otherwise be penalized by the metric for their low turnover. For example, in a 6 year crop sequence APPAAA, where A is an annual crop and P is a perennial, T₁ = 2, T₂ = 3, and T_p = 1 for a total turnover score of 6, whereas if P were an annual the total turnover would be 5. This adjustment makes the turnover term for consecutive perennials equivalent to the turnover if different annuals were grown in each year that the perennial is repeated.

The resulting metric:

$$\begin{align} {{\text{RCI}}} = \sqrt {n\frac{{{T_1} + {T_2} + {T_p}}}{2}} \end{align} \tag{ 2 }$$

yields a single value that can compare rotational complexity across any crop sequence of a given length and can be applied at fine scales with low computational costs (figure 2). The metric is able to cut through 'messy' data with unusual or unexpected crop rotations to sort sequences in terms of their expected potential to benefit soils (see representative RCI values in table 2).

Zoom In Zoom Out Reset image size

We identified publicly available datasets with a spatial component that might predict rotational complexity. Biophysical predictors included land capability, and magnitude and variability of precipitation during the growing season; policy predictors included distance from the nearest grain elevator and biofuel plant (table 1). We then compiled rasters from these datasets (figure S1 (available online at stacks.iop.org/ERL/16/054045/mmedia)).

As done in previous analyses [59–61], we used the NCCPI as a proxy for inherent land capability. The NCCPI combines soil properties (e.g. cation exchange capacity, bulk density, and slope) to form an index that shows agricultural land capacity, with submodels for corn/soy, cotton, and small grains [62]. In the present analysis, we use the highest NCCPI value out of the three submodels for each pixel.

2.2.1. Cropland data layer (CDL)

2.2.2. Field-level aggregation

To test for a relationship between RCI and predictive factors, all variables were centered and RCI was regressed against a set of covariate data (table 1) in a linear mixed model (LMM) including US state as a random effect to account for regional differences (see supplement). We included interactions for which we had a priori hypotheses (see table 3 for full list of terms included). The model was estimated using the R package 'lme4' [64].

Table 3. Regression coefficients. Coefficients from LMMs of RCI on various biophysical and social variables, with state as a random effect. Estimates are given from a naïve (non-bootstrapped) model, while confidence intervals are obtained from spatially blocked bootstrap regressions. After running 1000 bootstrapped models, the 2.5 percentile was used as the lower limit of the 95% confidence interval, while the 97.5 percentile was the upper limit. Interactions were chosen conservatively and a priori by author expectations about possible important interaction processes. Because variables were unscaled, magnitudes of coefficients are not directly comparable. Coefficients whose 95% confidence interval does not include zero are shown in bold.

Variable	Estimate	95% CI
Intercept	2.47 × 10−2	−4.43 × 10−2 to 6.99 × 10−2
Land capability	−2.84 × 10−1	−5.20 × 10−1 to −6.63 × 10−2
Grain elevator distance (km)	6.68 × 10−4	−2.21 × 10−3 to 3.63 × 10−3
Biofuel distance (km)	1.09 × 10−3	2.47 × 10−4 to 1.82 × 10−3
Mean rainfall (in)	−4.12 × 10−2	−5.93 × 10−2 to −1.07 × 10−2
Mean rainfall squared (in2)	3.82 × 10−3	−2.54 × 10−3 to 9.08 × 10−3
Rainfall variance (in2)	2.40 × 10−3	−3.04 × 10−3 to 4.12 × 10−3
Field size (ha)	−2.70 × 10−3	−2.99 × 10−3 to −1.93 × 10−3
Land capability × biofuel distance	−2.78 × 10−4	−2.86 × 10−3 to 3.81 × 10−3
Land capability × grain distance	−3.26 × 10−3	−1.85 × 10−2 to 7.21 × 10−3
Land capability × mean rainfall	−8.98 × 10−2	−1.73 × 10−1 to 1.02 × 10−1
Land capability × rainfall variance	−1.75 × 10−2	−2.48 × 10−2 to −4.08 × 10−3
Land capability × field size	6.27 × 10−3	1.31 × 10−3 to 7.81 × 10−3
Biofuel distance × field size	5.57 × 10−6	−5.67 × 10−6 to 1.28 × 10−5
Grain distance × field size	−1.31 × 10−5	−4.49 × 10−5 to 2.82 × 10−5
Mean rainfall × rainfall variance	1.86 × 10−3	−1.43 × 10−4 to 3.87 × 10−3
State random effect ($\sigma _{{\text{state}}}^2$)	4.52 × 10−2	2.52 × 10−2 to 8.59 × 10−2

Two model assumptions are violated in the above model, requiring updated estimates of the parameters' standard errors. First, because RCI is a derived statistic with an unusual domain, the index is not distributed according to a known distribution family and violates the assumption of normality in the residuals. Second, residuals showed high spatial autocorrelation at multiple scales (Global Moran's I = 0.23, p-value < 1 × 10⁻¹⁵, 20 nearest neighbors weights matrix) and with an unknown structure, necessitating a nonparametric approach. Both violations are likely to shrink standard errors of the estimated parameters, leading to overconfident estimates; to illustrate, in the case of spatial autocorrelation, if the explanatory variables are randomly located in relation to crop rotation, spatial autocorrelation in crop rotation would falsely inflate significance. We used nonparametric spatial block bootstrapping to correct for this overconfidence [65, 66]. An algorithm for sparsely distributed spatial data, derived by Lahiri 2018, was implemented in R (see supplement).

Spatial block bootstrapping involves iteratively resampling data in spatial blocks to mimic the generation of autocorrelated data. Choice of block size is nontrivial, and choosing the optimal block is an open question [67], but blocks should be larger than the scale at which autocorrelation operates. Using the R package 'gstat' [68, 69] to compute a variogram of the residuals generated by the naïve LMM, we determined that range (distance at which spatial autocorrelation falls off sufficiently) was 400 815 m. We used this as the dimension of each (square) spatial block (figure S2).

We repeated this bootstrap with a range of possible spatial block sizes and found that this inference on parameters was robust to the choice of block size (table S1).

RCI values calculated for corn-based rotations create the first map, to our knowledge, that quantifies field-scale rotational complexity across the US Midwest (figure 3).

Zoom In Zoom Out Reset image size

RCI values from 2012 to 2017 range from 0 to 5.2 (median = 2.2), and are positively skewed (figure 3(b)). Corn monoculture (i.e. corn every year for the 6 year period; RCI = 0) accounts for 4.5% of the study area and 3.3% of fields, suggesting that larger fields are more likely to be managed as monocultures (table 2). The mode RCI score (2.24) corresponds to a corn-soy rotation and dominates the region, covering over half of the study area. Two thirds of the area with this score was a CSCSCS or SCSCSC sequence, while the remaining third corresponds to other rotations that yield the same RCI (e.g. substituting soy for another crop, or a 3 year perennial followed by 3 years of corn).

RCI scores have statistically clear correlations with land capability, mean rainfall, distance to the nearest biofuel plant, and field size, as well as with several interactions between these variables (table 3; conditional R² = 0.14). Standard errors from the spatially blocked bootstrap were much larger than uncorrected naive confidence intervals, reflecting that accounting for spatial non-independence is necessary to estimate uncertainty of parameter estimates.

Rotational complexity decreased with NCCPI, a proxy for land capability. We find that land of higher inherent capability (flatter slopes, lower bulk density, etc) is more likely to be used for lower complexity rotations.

Rotational complexity decreased with average rainfall during the growing season. Fields with ample precipitation during the growing season are more likely to have simplified rotations.

Though the relationship between the proximity of the nearest grain elevator and a field's rotational complexity is not statistically clear (95% CI includes zero), RCI showed a clear increase with distance to the nearest biofuel plant. Fields that are closer to biofuel plants are therefore more likely to have simplified rotations.

Rotational complexity decreased with field size, with larger fields more likely to have simplified rotations.

Two of the interactions included in the model show statistically clear relationships. There is a positive interaction between land capability and field size, with higher quality land associated with decreasing RCI on small fields and slightly increasing RCI on large fields (figure 4(a)). The interaction between land capability and rainfall variance shows a negative effect on RCI, with highly variable rainfall accentuating land capability's impact on RCI (figure 4(b)).

Zoom In Zoom Out Reset image size

Interpretations of the relationship that each variable has with rotational complexity are shown in table 4. Though each change is associated with a small shift in average RCI across the region, these can represent massive shifts in regional land management.

RCI uses the sequence of cash crops on a given field as a proxy for crop rotation, and sorts these sequences into scores based on the sequence's complexity and potential for agro-ecosystem health. Because this metric has not been used in previous analyses, we verified RCI's validity through comparisons to previous estimates of rotational prevalence in the region. For example, two separate surveys of farmers in the US Midwest showed that between 24% and 46% report growing 'diversified rotations' [35, 70] which we consider to be an RCI of greater than 2.24 (i.e. corn-soybean). In the present study, 34% of fields had an RCI greater than 2.24. This and further comparisons of RCI to previous work (see supplement) show that RCI is capable of capturing previously-noted trends in the region [3, 34, 35, 63, 71].

The ability to analyze rotations at the field scale across the entire US Midwest allows us to ask how farmers optimize their rotations in complex economic and biophysical landscapes that include pressures to both simplify and diversify. Several biophysical and policy variables show statistically clear relationships with rotational complexity: high land capability, high rainfall during the growing season, and proximity to biofuel plants are all associated with rotational simplification.

Given policy incentives, farmers often find that 'corn on corn on dark dirt usually pencil out to be the way to go', with farmers growing corn year after year when high quality soil is available [35]. However, when that proverbial 'dark dirt' is not available, calculations are not so simple. If growing conditions are sufficiently poor (low land capability and low rainfall), these intensive corn systems may not be profitable, and farmers will have to rely more heavily on non-corn crops (or else inputs that eat into their profits) to maintain crop health and profitability in their fields.

We see this dynamic at play with land capability in the present analysis. Despite—or rather because of—the fact that more diverse rotations improve soils, the most degrading cropping systems counterintuitively tend to occur on the highest quality land. Highly capable lands can be farmed intensively without dipping into a production 'danger zone' in years with weather that is historically typical for the region, creating a pattern of land use that is likely to degrade these high quality lands in the long term and potentially jeopardize future yields, particularly in the face of climate change [25].

Recent analyses show that enhanced drought tolerance and resilience for crops is one of the key benefits of diverse crop rotations [25, 27, 72]. In the present analysis, mean rainfall during the growing season correlates positively with rotational simplification. Farmers may therefore be employing crop rotation in areas of low rainfall to achieve production levels that will keep a farm solvent, as was seen with rotational complexity increases in Nebraska during a drought period from 1999 to 2007 [73]. This trend is further accentuated by the negative interaction between land capability and rainfall variance in our analysis, where higher rainfall variability leads to even more diverse rotations on marginal lands.

Proximity to biofuel plants, the main policy indicator in our model, showed a statistically clear trend towards rotational simplification, likely due to increased economic profits. Local corn prices increase by $0.06–$0.12/bushel in the vicinity of a biofuel plant, amplifying incentives to grow corn more frequently [37]. Wang and Ortiz Bobea [39] were surprised not to find an impact of biofuel plant proximity on county-level frequencies of corn cropping in their own analysis, and the present analysis—done at a field rather than county scale—shows exactly this expected effect: corn-based rotations are simplified when in closer proximity to a biofuel plant.

In the current economic and policy landscape, farmers are pushed to simplify rotations through more frequent corn cropping, especially in proximity to biofuel plants, while marginal soils and low rainfall pull fields towards more diverse rotations.

RCI's ability to classify rotational complexity across large regions at the field scale and with low computational cost opens doors to future analyses that explore the interplay between localized landscape conditions, management choices, and agricultural, environmental, and economic outcomes. We see a strong potential to employ this metric not only in new regions, but in analyses that address how results from field experiments with crop rotation may scale up to regional levels [33].

We also note that the metric should be used with caution. For example, because RCI cannot recognize functional groups in crop sequences (e.g. legumes vs grains), it cannot capture the added benefits that diverse functional groups often add to a rotation. In addition, though RCI includes a perennial correction that avoids penalizing multiple consecutive years of perennials, the metric likely still underestimates the benefits of perennials in rotations. RCI is neutral to the soil benefits of annuals vs perennials, while in practice the year-round cover and crop species mixes (which are coded as a single crop in the CDL) that often accompany perennials may boost soil benefits beyond those of annuals [74]. Consecutive years of perennials are uncommon in our study area (e.g. less than 5% of studied crop area in Iowa), and we encourage caution before applying the metric to regions with a more substantial perennial presence.

We therefore recommend using RCI in studies that explore a wide range of cropping sequences where large differences in RCI are very likely to be meaningful, rather than as a tool to rank sequences that give similar scores. It is also important to note that, though the index can be applied to data of any sequence length, RCI values from different sequence lengths cannot be compared to each other; a rotation that results in a 2.2 from examining a 6 year sequence will not be a 2.2 when examining a 5 or 7 year sequence.

We also note that in using crop sequence as a proxy for crop rotation, RCI cannot fully capture the cyclical nature of true crop rotations. Because RCI examines a fixed number of years, it may 'split up' identical rotations in ways that give slightly different scores (for example, an 'AABB' rotation might appear as either AABBAA (RCI = 2.4) or ABBAAB (RCI = 2.6) in a 6 year sequence). As these discrepancies will decrease when longer sequences are considered, we recommend applying RCI to sequences that are as long or longer than the longest expected rotation in the study region.

We hope to see RCI used in future analyses that extend beyond the Midwest; however, regional and historic patterns of crop production likely influence farmers' rotational decisions and may render RCI scores calculated from disparate geographical regions difficult to interpret when called into direct comparison. We therefore see great promise in RCI as a rotational metric, and caution against applications that are overly narrow (e.g. comparing very similar RCI scores) and overly broad (e.g. comparing RCI scores across regions).

The time period chosen in this study, 2012–2017, coincides with the introduction of the Renewable Fuel Standard, or 'biofuel mandate', which took full effect in 2012. This policy mandates that 7.5 billion gallons of biofuel be blended with gasoline annually, and caused biofuel plants to open and local corn prices to soar across the US Midwest [75, 76]. Now in 2021, there is significant political pressure both to maintain the biofuel mandate in its current state and to relax the standards, and new exemptions to the mandate have already caused several biofuel plants to close in the region [77, 78]. Given the link between biofuel plant proximity and rotational complexity, our analysis suggests that these closures, if continued, would likely be associated with an increase in mean RCI in the US Midwest. Using our current model, simulations (n = 100) of randomly closing 20 of the 198 biofuel plants in the region lead to an increase of 0.003 in average RCI in the region, driven by greater distance to the nearest biofuel plant. In turn, increasing average RCI by 0.003 represents, for instance, the equivalent of 41 000 ha of cropland switching from corn-soy rotations to the most diverse rotation possible (six different crops grown across 6 years).

Rotational simplification near biofuel plants is a pertinent example of the influence that policy can have on farm management decisions and its landscape repercussions. Biofuel mandates are one of several policies, including crop insurance and research funding priorities [79], that currently maintain the profitability of corn production; however, these policies need not be the ones that define rotational landscapes, and increased funding for policies such as the USDA Conservation Stewardship Program could better align farmers' economic incentives with improved environmental health [80] (figure 1(b)).

When strong economic incentives encourage rotational simplification, our analysis suggests that it is more likely to occur on land with favorable biophysical conditions for corn growth. With our current policy structure, the highest quality lands in the US Midwest therefore become the most prone to degradation through intensive management.