A machine learning interpretation of the contribution of foliar fungicides to soybean yield in the north‐central United States

The surveyed, rainfed commercial soybean fields were spread across the U.S. north-central region (Supplementary Fig. S1 online) with a latitudinal gradient evident for maturity group (MG). The number of fields (n) was distributed evenly across the three years (2014: n = 812, 2015: n = 960, 2016: n = 966). Among the 2738 fields, 833 (or 30.4%) were sprayed with foliar fungicides. Out of the 833 fields sprayed with foliar fungicides, 623 (74.8%) had also been sprayed with foliar insecticides.

A t-test estimate of the yield difference between all fields sprayed with foliar fungicides and those which were not was 0.46 t/ha (95% confidence interval [CI] of 0.39 to 0.52 t/ha). When t-tests were applied to fields within TEDs (the 12 TEDs with the most fields), half of the 95% CIs included zero, indicative of possibly no yield increase due to foliar fungicides over unsprayed fields in those TEDs (Supplementary Fig. S2 online). A linear mixed model with random slopes and intercepts for the fungicide effect within TEDs returned an estimated yield gain of 0.33 t/ha due to foliar fungicide use. A simpler model without random slopes for foliar fungicide was a worse fit to the data. Together these basic tests were indicative of heterogenous effects concerning foliar fungicides and yield gain, implying other global (regional) and local (field specific) conditions may be involved as factors.

A tuned random forest (RF) model fitted to the entire dataset (all 2738 observations) overpredicted soybean yield at low actual yields, and underpredicted at the high-yield end (Supplementary Fig. S3 online). However, as 99% of the residual values were less than or equal to |0.25 t/ha| which corresponded to less than 7% of the average yield, we proceeded with the interpretation of the fit RF model. The mean predicted soybean yield (global average) was 3.79 t/ha (minimum = 1.13 t/ha, maximum = 6.02 t/ha, standard deviation = 0.81 t/ha, root mean squared error between the observed and predicted yields = 0.1 t/ha).

At the global model level, location (latitude; a surrogate for other unmeasured variables) and sowing date (day of year from Jan 01) were the two variables most associated with yield (Fig. 1), consistent with the central importance of early planting to soybean yield^5,13. Soil-related properties (pH and organic matter content of the topsoil) were also associated with yield (Fig. 1). Management-related variables such as foliar fungicide, insecticide and herbicide applications were of intermediate importance, and other management variables (row spacing, seed treatments, starter fertilizer) were on the lower end of the importance spectrum in predicting soybean yield (Fig. 1). Insecticide and fungicide seed treatments were poorly associated with soybean yield increases as has been previously shown^8,40. The relatively lower importance of row spacing is consistent with previous analyses of this variable from soybean grower data⁶. The dataset we analyzed did not contain enough observations to include artificial drainage as a variable, which has been shown to influence soybean yield, presumably by allowing earlier sowing¹⁴.

The strongest pairwise interactions included that between sowing date and latitude. Delayed sowing at higher latitudes decreased yield by about 1 t/ha relative to the highest yielding fields sown early in the more southerly locations (Supplementary Fig. S4 online). Further examination of the interactions showed that the yield difference between sprayed and unsprayed fields increased with later sowing, indicative of a greater fungicide benefit in later-planted fields (Fig. 2). This would seem to conflict with the results of a recent meta-analysis in which soybean yields responded better when foliar fungicides were applied to early-planted fields²⁷, but in that study there was also the confounding effect of higher-than-average rainfall between sowing and the R3 growth stage. With respect to latitude, the global difference in yield between sprayed and unsprayed fields decreased as one moved further north (Fig. 2), suggesting that foliar fungicides were of more benefit when applied to the more southerly located fields, which do tend to experience more or prolonged conditions conducive to foliar diseases than the northern fields^22,24.

Focusing on model interpretation at the local level, we examined the Shapley φ values (see the “Methods” section for more information) associated with foliar fungicide applications for different subsets (s) and cohorts (c) of fields within the data (see Supplementary Table S1 online). The 1st subset (s₁) was comprised of the 20 highest-yielding fields among those sprayed with foliar fungicides (s₁c₁) and the 20 highest-yielding fields among those which were not sprayed (s₁c₂) in each of the 12 technology extrapolation domains (TEDs) in the data matrix with adequate numbers of fields for comparisons (see also Supplementary Table S2 online; Supplementary Fig. S5 online maps the field locations within these 12 TEDs). A TED is a region (not necessarily spatially contiguous) with similar biophysical properties⁴¹. Predicted yields within these cohorts were mainly above the global average of 3.79 t/ha, except in TED 602303 (Fig. 3), which corresponded to fields in North Dakota (Supplementary Fig. S5). In most cases Shapley φ values for foliar fungicide use exhibited a positive contribution to the yield above the global average. If these cohorts of fields represented high-yielding environments within each TED, then foliar fungicide sprays contributed positively up to 0.3 t/ha in the yield increase above the global average in s₁c₁. However, among high-yielding fields in s₁c₂, the penalty for not spraying was less than 0.1 t/ha. This finding supports the contention that fungicide sprays are most worthwhile in high-yielding environments. Supplementary Fig. S6 online complements Fig. 3 by summarizing the Shapley φ values in another visual format. The overall mean predicted yield for the unsprayed (s₁c₂) fields was slightly higher (by 0.1 t/ha) than that for the sprayed (s₁c₁) fields (Supplementary Fig. S6 online). This difference may have been driven by the higher variability in yields among the two cohorts (particularly for TEDs 403603, 602303, 403703, and 303603), or underlying differences in other management factors. Also, the number of sprayed fields in each of these four TEDs was at the target sampling boundary of 20 fields per TED (Supplementary Table S2 online).

The Shapley φ values for fungicide use were well-separated among the four cohorts of fields of s₂ (Fig. 4, Supplementary Table S1 online). The fields within s₂ were selected across the entire dataset and not by TED membership. The lowest-yielding fields (s₂c₂ & s₂c₄) were all below the global yield average, whereas the converse was true of the highest-yielding fields (s₂c₁ & s₂c₃). Among the lowest-yielding fields, foliar fungicides were mainly associated with a positive, but less than 0.2 t/ha, effect on yield (s₂c₂), and other factors were responsible for dropping a field’s yield to below the global average. Amongst the highest-yielding fields (s₂c₁), foliar fungicides were associated with between 0.15 and 0.35 t/ha of the yield above the global average. These Shapley φ values for the contribution of foliar fungicides are consistent with estimates of the yield response to foliar fungicides from a meta-analytic perspective²⁷. Given that the individual yields in s₂c₁ & s₂c₃ were 1 to 2 t/ha above the global average, other location-driven factors such as early sowing (Fig. 1) were the larger drivers of yield in these cases. However, there was only a negligible or small (< 0.1 t/ha) penalty for not using foliar fungicides in high-yield situations (s₂c₃; see also Supplementary Fig. S7 online).

There was some overlap in the fields of s₂ and s₃ [where s₃ consisted of fields within the 90th percentile for yield among sprayed fields (s₃c₁); and the 90th percentile for yield among the unsprayed fields (s₃c₂) in the dataset], at least where high-yielding fields were concerned. All fields in s₃ had predicted yields that were above the global average (Fig. 5). Yield distributions of the two cohorts within s₃ were similar, with the cohorts having near-identical mean yields. Foliar fungicides contributed to between 0.1 t/ha and 0.35 t/ha to the yield increase above the global average, while the penalty (if there was one) for not using foliar fungicides was mainly confined to less than 0.05 t/ha, indicating that among the fields of s₃c₂ spraying was unnecessary (otherwise the penalty would have been larger). Overlaying the estimated φ values for fungicide use with MG, sowing date and growing degree days showed that these high-yielding fields were mainly in MG II and III, that the fields tended to be planted early, and were restricted to GDD groups 03 and 04 (Fig. 5), the latter factor being highly aligned with latitude. A formal comparison of the Shapley φ values across cohorts was not attempted because they potentially differed in their underlying variables despite similar yield distributions within the lowest- or highest-yield cohorts.

Intuitively, one may have expected the yield increase due to foliar fungicides to be about the same magnitude (about 0.1 t/ha) as the yield penalty associated with not using fungicides. The larger yield gain versus the penalty may be due to synergistic interactions of foliar fungicides with other management factors. For example, foliar insecticides are likely to be applied along with foliar fungicides; conversely, fields that were not sprayed with foliar fungicides were unlikely to be sprayed with insecticides as well. Therefore, in subset 4 (s₄), we examined the Shapley φ values associated with fungicide use among all 210 fields in the data matrix which had been sprayed with foliar fungicides but not with foliar insecticides (s₄c₁), and compared them to the Shapley φ values for foliar fungicide use among another cohort of 210 fields (s₄c₂) which had been sprayed with both foliar fungicides and insecticides, where the fields of s₄c₂ were sampled to match the range of reported yields in s₄c₁. There was no discernable separation of the Shapley φ values between cohorts s₄c₁ and s₄c₂ (Supplementary Fig. S8 online), and the φ values were consistent with what had been observed with the other subsets of fields.

A partial economic analysis estimated the net realized profit associated with foliar fungicide use on the respective cohorts within subsets of fields. The profitability of foliar fungicides in the fields of s₁c₁ (20 highest-yielding sprayed fields within the 12 TEDs with the most fields in the dataset) is shown in Fig. 6. It should be noted that the soybean price on which Fig. 6 is predicated reflects the high prices being experienced currently (as of Spring 2021), which are at their highest levels in at least the last five years. Assuming a price of US$576.30/t, fungicides were overwhelmingly profitable in all but four TEDs (403603, 602303, 403703, 303603) in which the average return (with respect to fungicide use) was less than US$7.50/ha. For these four TEDs, confidence intervals for the mean financial return per ha after accounting for fungicide costs indicated returns could be negative (loss), zero, or up to US$26.50/ha, depending on the individual field (Supplementary Figure S9 online). Considering these were the highest-yielding fields within TEDs, there was the risk of losing money on fungicide sprays in these four TEDs. Obviously, environment mattered (compare with Supplementary Figures S5 and S6 online), and with the four TEDs listed above the most noticeable feature was their higher-latitude locations relative to fields in other TEDs. Among other things, higher latitude is associated with cooler weather and shorter accumulation of GDD. Underlying yield potential factors (early sowing, PAWR, GDD, AI) contributed to higher predicted yield. Higher-yielding environments were more likely to also realize a larger contribution of foliar fungicides to yield above the global average (Supplementary Figure S6 online), thereby leading to the profitability of spraying.

The financial return on spraying the fields in s₂c₂ (100 lowest-yielding fungicide-sprayed fields) was negative, except in a few individual cases (Supplementary Figure S10 online). The mean net return due to foliar fungicides for s₂c₁ (100 highest-yielding fungicide-sprayed fields) was US$74.63/ha (95% CI US$69.02 to US$80.66 per ha), whereas for s₂c₂ the return was -US$26.24/ha (95% CI -US$33.63 to -US$19.71 per ha).

Considering the two cohorts of s₃ (unsprayed and sprayed fields in the 90th percentile for yield), there was a small financial penalty to not using foliar fungicides in high-yield environments. Not spraying high-yield fields (s₃c₂) was associated with a mean loss of -US$10.17/ha (95% CI -US$12.50 to -US$7.83 per ha). Yet, spraying high-yield fields (s₃c₁) was associated with a mean gain of US$65.60 (95% CI US$61.28 to US$70.38 per ha). The trends were consistent when other soybean price points were assumed for any of the subsets examined (Supplementary Figures S11, S12, S13 online).

The soybean price required to at least break even on a (fixed) fungicide investment cost of US$61.90/ha was a nonlinear function of φ. At a realized Shapley φ value of 0.1 t/ha in response to foliar fungicides, soybean price would have to be at least US$619.00/t to recover the costs of fungicides and their application, dropping to US$309.50/t, US$206.33/t, and US$154.75/t for Shapley φ value of 0.2 t/ha, 0.3 t/ha and 0.4 t/ha, respectively.

The percentage of U.S. soybean hectarage treated with foliar fungicides rose from 1 to 11% between 2004 and 2015⁴², which is a yearly increase of 0.91%. Assuming the average gain of 0.221 t/ha due to foliar fungicides among sprayed fields in the 90th percentile for yield (s₂c₁), we estimated a yield gain of 2 kg ha⁻¹ year⁻¹ attributed to the adoption of foliar fungicide (221 × 0.91/100). This translated to 6% of the estimated annual yield gain in U.S. soybean (33 kg ha⁻¹ year⁻¹) attributable to foliar fungicide use in high-yield environments.

As foliar disease data were not available, we can only say that a decision was made to use foliar fungicides in about one-third of the fields, but cannot say why growers chose to spray, which could be any one of (or a combination) of cost effectiveness, perceived benefit (disease control or plant health effects) or forecast disease risk. Whatever the reason, the estimated yield gains (above the global average) attributed to foliar fungicides made spraying profitable under several soybean price scenarios, but the yield potential environment is an important consideration as highlighted by a loss on fungicide investment in some TEDs. Our finding that fungicide profitability was not universal may account for some of the discrepancies among field trials mentioned in the Introduction.

We do emphasize that foliar fungicides should not be applied indiscriminately, divorced from disease scouting or forecasting, integrated pest management and environmental principles. The price to be paid in terms of environmental damage⁴³ and loss of product efficacy due to the evolution of fungicide resistance within foliar pathogen populations^44,45 should be weighed against the yield penalty associated with not using foliar fungicides in high-yield environments. For the unsprayed fields in the 90th percentile for yield (s₃c₂), the average penalty associated with not spraying was 17.7 kg/ha, which works out to be US$10.62/ha at a high price of US$600/t.

Source: Ecology - nature.com