Satellites reveal a small positive yield effect from conservation tillage across the US Corn Belt

The US Corn Belt encompasses approximately 1 million km² across 12 states in the midwestern United States [40]. It is characterized by high-yielding commercial agriculture predominantly in maize-soy rotation, contributing over one third of global production for these crops [41]. Here, we focused on a 9-state region for maize (figure 1(a)) and 3 states (Indiana, Illinois, and Iowa) for soybeans due to yield map availability for each crop (section 2.2). This region has primarily hot-summer to warm-summer humid continental climates (Koeppengeiger classes Dfa and Dfb) [42], and most fields do not receive supplemental irrigation. According to the US Agricultural Census [43], conservation tillage covered 50% (∼412 000 km²) of total cropland area in the 9-state region in 2017, a 17% increase since the previous 2012 Census. Cover cropping, a complementary practice promoted along with reduced tillage and crop rotation as the three pillars of 'conservation agriculture,' is not nearly as prevalent, covering only 3.4% of total cropland area in the 2017 Census. Still, this represents a 75% increase since 2012 [43]. Increasing adoption of these soil conservation practices has reduced average erosion rates in the United States ∼35% between 1982 and 2007, but rates remain above natural soil production [44].

Zoom In Zoom Out Reset image size

We used a previously published gridded dataset of annual tillage practices for the north central US from 2005 to 2016 by Azzari et al [35] to identify locations practicing conservation or conventional tillage at 30 m resolution (figure 1(a)). Briefly, these maps were generated by applying a random forest classifier trained on ground truth data from 5866 soybean fields to Landsat satellite imagery. Because the ground truth data was limited to soybean fields, the classification was applied only to pixels identified as soybeans based on annual crop type maps from the US National Agricultural Statistics Service (NASS) [45]. These soybean-based tillage maps achieve fairly complete coverage of the study region every two years due to dominant maize-soybean rotations. Here, we assume that the same tillage method was practiced during subsequent maize years. It is likely this assumption is not universally valid, since partial adoption characterizes over half of conservation tillage practitioners, with approximately 11% of farmers in this region adopting tillage practices by crop type [46]. Similarly challenging for inference, this product has an overall accuracy of 79%, with 84% and 72% of validation points correctly classified for conservation and conventional tillage, respectively. We mitigate these potential sources of error through additional data filtering criteria (section 2.3) and note that, given our findings (section 3.2), the possible inclusion of a small percentage of misclassified maize fields implies that our estimates of yield differences between conservation and conventional tillage systems are likely conservative.

Previously published yield maps for maize and soybean were produced using the Scalable Crop Yield Mapper (SCYM) [33], a satellite-based approach with a demonstrated ability to detect impacts from management practices [47, 48]. This approach has two main steps. First, statistical models predicting yields from crop phenology and climate covariates are derived from regionally parameterized crop models. Second, these statistical models are applied to satellite imagery and gridded climate datasets based on crop type maps, generating a yield estimate for each pixel. Maize yield maps were produced with Landsat satellite imagery for nine Corn Belt states from 2008 to 2015 by Jin et al [34]. We used the algorithm described therein to extend maize yield maps through 2017, with overall county-level agreement at r² = 0.78 (RMSE = 1.2 t/ha) compared with NASS statistics. Soybean yield maps were produced by Lobell and Azzari [36] for the states of Indiana, Illinois, and Iowa from 2000 to 2015 with similar agreement to county yield statistics (r² = 0.74, RMSE = 0.16 t/ha). Figure S1 (figure S1 is available online at stacks.iop.org/ERL/14/124038/mmedia) shows mean yields across space for each crop during the study period.

As noted above, the tillage map classification performs moderately well but contains some errors that could add noise to our analyses. To guard against spurious classifications, we restricted the tillage maps in the following ways: (1) we required pixels to have 6 observations during the 12 year dataset, with at least one observation before 2008 and after 2014, to ensure a dense time series of observations spanning the data record; (2) from these, we identified long-term tillage management regimes based on pixels with constant tillage status in all observations, indicating at least a decade of conservation practices and likely increasing the probability of sustained adoption in maize years; (3) we identified 'single-switch' pixels which switched tillage status one time between 2009 and 2014 to examine the impact of new tillage regimes; and (4) pixels needed to be part of a coherent pixel group of the same tillage classification and, for single-switch pixels, the same year (see Text S1). All remaining sampling and analyses occurred at this 'field entity' level. We then used a data-driven delineation of climate-soil domains [26] to sample up to 500 fields per tillage status within each domain for each year, resulting in 144 127 and 117 757 maize field-years and 92 037 and 100 222 soybean fields-years for conventional and conservation tillage, respectively.

For each field, we extracted the median yield from SCYM maps. We removed fields with outlier yield values below the 0.01% and above the 99.99% from both maize and soybean datasets. We then extracted median field values for a suite of environmental covariables defining both static field properties and annually varying weather and soil moisture. For static field properties, we obtained 1981–2010 climate normals from PRISM [49, 50], calculated field slope from the USGS National Elevation Dataset [51], and extracted soil properties for the top one meter from the SSURGO soil database [52]. Annual monthly and seasonal weather summaries were extracted from GRIDMET ∼4 km meteorological dataset [53]. Annual monthly modeled soil moisture and climate water deficit were extracted from the TerraClimate ∼4 km climatic water balance dataset [54]. Table S1 provides a list of all variables considered and their data sources. All data were accessed and processed in Google Earth Engine [55] with the exception of the soil data, which we acquired through SSURGO.

To quantify conservation tillage's impact on crop yields, we used causal forests, an recent adaptation of the classic random forest algorithm [56] for statistical inference on causal effects, particularly when heterogeneity is present [38, 39]. Broadly, causal forests act as an adaptive kernel method [39]; in our application, it uses each field's closest neighbors in covariate space to generate a counterfactual yield estimate under the alternative tillage practice. Causal forests generate mathematically valid confidence intervals while leveraging the ability of random forests to handle many covariates and nonlinear interactions without overfitting or requiring explicit model specification [38, 39, 57, 58]. Recent applications demonstrate better performance than conventional econometric methods for detecting and quantifying heterogenous treatment effects [59, 60].

Causal forests are also designed for observational datasets. Because treatments are not randomly assigned, an observational analysis could be confounded if fields that have higher (or lower) yields also tend to adopt conservation tillage at higher rates. Causal forests addresses these biases with a 'doubly robust' treatment estimation method (augmented inverse-propensity weighted estimation [61]) which combines both treatment propensity weighting [62] and regression adjustment to reduce sensitivity to misspecification in either model [39, 63].

2.4.1. Analysis of long-term conservation tillage

Here, we used the 'grf' package [64] in R [65] to implement causal forests separately for maize and soybean fields with long-term tillage practices (section 2.3). We designated 'conservation tillage' as the treatment variable, 'conventional tillage' as the control, and crop yield as the outcome. First, we used the full set of static covariates describing field slope, soil properties, and climate normals (tables 1 and 2) to estimate treatment propensity using 2000 trees and default function settings. The propensity model performed well, indicated by close agreement between propensity scores versus treatment status (figure S3). To examine biophysical factors typically associated with conservation tillage, we used the larger 9-state maize domain. We inferred variable importance from the number of times each covariate was used to split the individual trees, although it should be noted that correlations among variables can skew these metrics [66].

Table 1.  Variables used in causal forests analysis: Maize. Variables are ordered by the proportion of splits on each variable within the ensemble of decision trees that make up each forest (high to low), which provides a rough approximation of variable importance. VPD = vapor pressure deficit; AWC = available water content. Table S1 provides data source information.

Treatment propensity (30 yr climate normals)	Regression adustment/Expected yield outcome	Treatment effect
Slope	July mean max temp	Slope
May precip	May mean max temp	Soil clay content
April mean temp	Year	Soil sand content
April precip	July climate water deficit	June precip
July VPD	Solar radiation (June–August)	April soil moisture
June precip	August mean max temp	May mean min temp
July precip	May mean min temp	Early season precip
June VPD	June precip	August soil moisture
Soil sand content	Early season precip	Soil AWC
June mean temp	August soil moisture	June mean min temp
Soil silt content	June mean min temp	May mean max temp
Soil AWC	Growing Degree Days	June climate water deficit
May mean temp	May precip	July VPD
August mean temp	July precip	July mean max temp
July mean temp	August mean min temp	May precip
Soil clay content	Soil clay content	Growing Degree Days
Soil ksat	Soil AWC	April precip
	Soil sand content	Solar radiation (June–August)
		August mean min temp
		August mean max temp
		July precip
		April mean temperature
		July climate water deficit
		Year

To meet the assumption of overlap within the causal forests framework, which requires that treatment and control samples occupy similar covariate space to provide appropriate neighbors for comparison, we then removed samples with propensity scores below 0.05 or above 0.95. This produced a final dataset of 70 404 and 88 220 (maize) and 51 215 and 68 334 (soybeans) unique field-year observations for conservation and conventional tillage, respectively. Figures S4 and S5 provide the spatial distribution of field observations before and after this propensity filter. We then specified the regression adjustment portion of the doubly robust estimator (see Text S2 and figure S2). Next, we used all covariables selected for this regression model and the most important variables in the propensity model (tables 1 and 2) to estimate the treatment effects of conservation tillage using the 'causal_forest' function in grf with 2000 trees and default parameters.

Table 2.  Variables used in causal forests analysis: Soybeans. Variables are ordered by the proportion of splits on each variable within the ensemble of decision trees that make up each forest (high to low), which provides a rough approximation of variable importance. VPD = vapor pressure deficit; AWC = available water content. Table S1 provides data source information.

Treatment propensity (30 yr climate normals)	Regression adjustment/Expected yield outcome	Treatment effect
Slope	Aridity (June–August)	May soil moisture
May precip	Previously Soy	April soil moisture
April mean temp	July climate water deficit	Slope
Soil silt content	Solar radiation (June–August)	July soil moisture
June VPD	June precip	Soil silt content
April precip	Growing Degree Days	July precip
Soil clay content	May soil moistre	Aridity (June–August)
July precip	June climate water deficit	Soil clay content
June precip	May precip	Soil sand content
May mean temp	Jan–April precipitation	June precip
August mean temp	July precip	August mean min temp
July mean temp	July soil moisture	Growing Degree Days
Soil AWC	Soil sand content	Soil AWC
June mean temp	April precip	July VPD
July VPD	June–August precip	Early season precip
Soil sand content	Soil clay content	Solar radiation (June–August)
Soil ksat	August mean min temp	April mean temp
	Year	May precip
		Growing season precip
		Previously Soy
		April precip
		June mean min temp
		Year
		June climate water deficit
		July climate water deficit

To investigate heterogeneity in treatment effects, we first tested for significant heterogeneity using the 'test_calibration' function in grf. We then summarized covariate values for subpopulations of observations based on their predicted treatment effects. Figures 2(a) and 3(a) show the distribution of field samples among these subpopulation bins. We then identified covariates with stronger influences on the yield outcomes of conservation tillage based on covariate distributions within these subpopulations and informed by variable importance rankings for the causal forests (tables 1 and 2).

Zoom In Zoom Out Reset image size

2.4.2. Analysis of fields following initial tillage conversion

2.4.3. Confounders and omitted variables

Overall, we found that long-term conservation tillage occurred across a wide range of environmental conditions, as indicated by treatment probabilities above 25% across most covariable values (figure 1(b)). This is consistent with past difficulties identifying universal variables to explain adoption [71, 72]. Field slope ranked highest in importance (table 1), likely explained by greater benefits for erosion relative to flatter fields and policies targeting highly erodible areas [18]. Higher early season temperature, early season precipitation, and July vapor pressure deficits also increased probability of adoption (figure 1(b), table 1), consistent with findings that conservation tillage is often more prominent in warmer, arid conditions [16, 17, 73] and can enhance water infiltration rates [1]. Both soil silt and available water content were positively related to conservation tillage while sand content was negatively correlated (figure 1(b)), although soil variables did not rank high in variable importance.

For maize, we found an overall 3.3% yield benefit from conservation tillage (average treatment effect; 95% Confidence Interval, CI = [3.1%, 3.4%]) across fields with long-term tillage practices from 2008 to 2017. This translated to an average yield boost of 0.36 t/ha (CI = [0.34, 0.37]). For soybeans, we found a smaller yield benefit of 0.74% (CI = 0.53%, 0.95%) or 0.024 t/ha (CI = 0.017, 0.031) from 2005 to 2015 within Indiana, Illinois, and Iowa. Generally, these effects are smaller than typical year-to-year yield variability in this region (figure S2) and similar in magnitude to previous plot level work [20, 21, 27, 74, 75]. We did, however, find more evidence for positive yield impacts across the region than the majority of existing literature [16]. This may reflect ongoing technology improvements for conservation tillage implementation [29] or additional insight afforded through our methodology, which allows the inclusion of a large range of covariates and leverages thousands of fields across a wide region.

Tests for treatment effect heterogeneity were significant for both maize (p < 0.0001) and soybeans (p < 0.0001), indicating treatment effects are moderated by the weather, soil, and slope covariates used. For maize, the 5th–95th percentiles of these conditional average treatment effects (CATEs) ranged from −1.3% to 8.1%; for soybeans, they ranged from −4.7% to 5.8%.

Zoom In Zoom Out Reset image size

To understand how this heterogeneity manifested across the Corn Belt, we mapped the mean CATE for all field-years on a 5 km² regular grid by crop type (figures 2(a) and 3(a)). For both maize and soybeans, negative impacts from conservation tillage were most pronounced in northwestern Iowa and from southeast Iowa into western Illinois. Conservation tillage largely improved yields from eastern Illinois through Indiana. For maize, Ohio, South Dakota, and the outer regions of the Corn Belt displayed strong positive effects. Notably, we found that conservation tillage had a largely positive effect on maize yields in the northern Corn Belt, where it has historically been more limited in practice (figure 1(a)). While it is possible this effect is driven by early adopters inordinately adept at managing their fields, recent studies in Minnesota [29, 76], New York [12], and Canada [32] provide increasing evidence for conservation tillage interest and feasibility in these more northerly latitudes.

To understand the underlying biophysical features driving these patterns, we explored conditional treatment effects by field attributes for both maize (figure 2(b)) and soybeans (figure 3(b)). Overall, the soil water balance and seasonal temperatures seem to drive much of the heterogeneity observed. For example, maize and soybean yield benefits were greater than average when baseline late-season (July–August) soil moisture was low, suggesting higher differential success in arid conditions likely due to improved soil water holding capacity. Similarly, soybean field-years with positive yield effects tended to have lower baseline soil available water content from static soil maps, indicating potential improvement in water capacities on these fields from conservation tillage.

In addition to these arid conditions generally thought to benefit from conservation tillage [16], we also found evidence that conservation tillage can improve yields under wet conditions. Soybean field-years with positive yield impacts had higher median July precipitation (figure 3(b)). Similarly, maize field-years with the greatest treatment effects experienced higher median early season precipitation (figure 2(b)). Together, this suggests improved water infiltration on fields under conservation tillage.

On the other hand, we also found evidence that very wet early season soils (April–May) can reduce conservation tillage benefits in both crops. Conventional tillage helps dry water-logged soils [77], often enabling earlier planting dates and thus better yields. Interestingly, conservation tillage performed worse for maize when mean May minimum temperature was higher (figure 2(b)). Although higher temperatures should help dry soils near planting, higher May temperatures could also facilitate weed growth that can compete with maize emergence or, combined with residue cover, increase disease pressure by fostering disease organisms.

Although higher field slopes increased the likelihood of conservation tillage (figure 1(b)), yields were not better on higher slopes for either crop (figures 2(b), 3(b)), possibly because impacts from soil erosion operate on centennial time scales not yet manifested here. However, we were unable to compare fields with slopes higher than 3 degrees due to lack of overlap between tillage types, since high sloped fields had high probabilities of treatment (figures 1(b), section 2.4.1).

There is strong evidence that any benefits from conservation tillage can be absent upon initial implementation and accrue over time as soil health and management improves [16]. For maize and soybeans, we found an overall positive yield effect of 0.29% and 0.033%, respectively, for each additional year under conservation tillage when considering fields between 1 and 8 years since adoption. Although still positive, these effects are an order of magnitude smaller than fields with long-term conservation tillage (figure 4). These numbers imply that, on average, the full yield benefit of long-term conservation tillage is achieved after 11 years for maize and 22 years for soybeans. Interestingly, we found similar (maize) or greater (soybeans) yield improvements when analyzing fields that switched from conservation tillage to conventional tillage (maize: 0.26%; soy: 0.61%). This suggests that management challenges persist for conservation tillage, likely related to weed control or timing of planting. Indeed, a previous analysis found that in the US circa 2012, less than half of farmers reporting 'no-till' methods practiced them continuously during the previous four years [18]. Because soil benefits are greatest under sustained conservation tillage, there is still a need for improved understanding of these management decisions and challenges.

Zoom In Zoom Out Reset image size