Telling Apart the Bad from the Good Guys Behind the Spraying Mist

14 Residues of pesticides not allowed in organic farming are often found in organic food. A large 15 number of samples are being tested by organic certifiers, but the sampling methods often do 16 not allow to determine if such residues stem from prohibited pesticide use by organic farm- 17 ers, from mixing organic with conventional products, from short-range spray-drift from neigh- 18 bour farms, from the ubiquitous presence of such substances due to long-distance drift, or 19 from other sources of contamination. Eight case studies from different crops and countries 20 are used to demonstrate that sampling at different distances from possible sources of short- 21 distance drift allows in most cases to differentiate deliberate pesticide application by the or- 22 ganic farmer from drift. Datasets from 67 banana farms in Ecuador, where aerial fungicide 23 spraying leads to a heavy drift problem, were subjected to statistical analysis. A linear discri- 24 minant function including four variables was identified for distinguishing under these condi- 25 tions application from drift, with an accuracy of 93.3%.

rows. While the figures for conventional products remain in the same range in the process from whole-50 sale to retail (blue rectangle to the right), the residues for organic products are very substantially re-51 duced during this process (green trapezium). As a result, residues at retail level are 100 and more 52 times lower in organic than in conventional products (trapezium at the bottom). This shows that the 53 process represented by the red arrows works fairly wellwhich is not always the case for the investi-54 gation of the origin of such residues, as represented by the yellow arrows.

55
Unfortunately, this good news for consumers with respect to objective (b) does not always 56 mean that objectives (a) and (c) are also met. With the steady growth of the organic market 57 and globalisation of supply chains, fraud has also grown. 4,5 Pesticide residues in organic 58 products can be the result of fraudulent spraying by farmers, commingling organic with con-59 ventional products, selling conventional products as organic, spray-drift, or different (avoida-60 ble or unavoidable) sources of contamination along the supply chain. 61

62
Over the past decades, a distinction has been made between short distance primary spray-63 drift during the application, and long distance secondary spray-drift occurring after the ap-64 plication. 6 The latter was attributed to evaporation and considered to play a role only for pes-65 ticides with high vapour pressure. 7 On the one hand, recent studies have shown that evapo-66 ration and long-distance transport can already play a role during, not only after application. 8 67 On the other hand, long-distance transport has been found to be linked not only to evapora-68 tion. Pesticides adherent to dust from wind erosion can contaminate large areas. 9 In the pre-69 sent context, we use the terms short-range and long-range drift, instead of primary and 70 secondary drift ( Figure 2). 71

72
Long-range drift is so far poorly understood, can lead to (normally very low) residues at dis-73 tances as far as thousands of km, 10 and happens in the form of vapour or molecules adhered 74 to dust. The main factors influencing long-range drift are vapour pressure of the pesticide, 75 capacity of adherence to dust, incidence of wind erosion, and temperature inversion in the 76 atmosphere. 7 Long-range pesticide drift has recently received more attention. 11,12,13,14, 15 Ex-77 amples have been used in the context of organic certification for supporting the argument of 80  have been quoted to demonstrate the ubiquity of pesticides. 17 None of these case studies, 92 however, provides solid evidence for the assumption that long-distance transport of pesti-93 cides leads to residues in organic food above the level of, say, 0.01 to 0.03 mg/kg. The prob-94 lem of the herbicides pendimethalin and prosulfocarb being subject to long-distance drift be-95 cause of their high vapour pressure, has been known for a long time, 18 but this phenomenon 96 cannot be extrapolated to other substances. Even for these herbicides, there is no evidence 97 that residues at larger distances could be above the indicated levels. On an average of 15 98 vegetation samples from nature reserves in Germany, 0.009 mg/kg pendimethalin and 0.004 99 mg/kg prosulfocarb were found. 19 Exceptions may exist, e.g., when pesticide applications are 100 followed by heavy wind erosion, as seems to be the case in some of the North American 101 wheat growing areas, where glyphosate is used for cereal desiccation shortly before harvest. 134 knowledge about short-range drift as a tool for assessing farmers' compliance with organic 135 production rules. The dynamics of short-range spray-drift have been widely studied in the 136 context of preventing liability problems due to herbicide damage, contamination of water bod-137 ies and natural habitats, and direct risks for human settlements. 10,21,22,23,25,26,27 Pesticide de-138 posit decreases exponentially with increasing distance from the field, on which the substance 139 is applied. With a tractor boom sprayer, deposit at 25 m distance is expected to be only 1% 140 of that in the target field. While distances are greater for air-blast or aerial spraying, the basic 141 principle of exponential decrease is the same (Figure 2 and Supplementary Fig. 1).

Certifiers' Testing Strategies
Both the EU Regulation on organic farming and the US National Organic Program (NOP) re-144 quire certification bodies (CBs) to take samples from at least 5% of their clients every year. A 145 large amount of data is being generated through this mechanism, but the sampling proce-146 dures and interpretation of results often do not allow to derive clear results. 147 Testing at different points along the organic supply chain could be an excellent tool for de-148 tecting non-compliance with organic production rules. The idea behind this is depicted in Fi-149 gure 1. The filter process as such, and the exclusion of contaminated batches from the or-150 ganic market, as represented by the red arrows, often work well. Thus, there are significantly 151 lower average amounts of residues after undergoing this filtering process across the supply 152 chain. The fact that residues in organic produce reported from the wholesale and processing 153 levels are massively higher than those reported from retail samples (33 times higher for veg-154 etables and 7 times higher for fruits) while they vary little for conventional produce, shows 155 that market actors often remove problematic produce by declaring it conventional. However, 156 the information about these switches is not always reaching the CBs, thus impeding the in-157 vestigation of the origin of residues and the exclusion of excluding fraudulent actors from the 158 market (yellow arrows in Figure 1). 159 Not only for market actors, however, but also for many CBs, the purpose of sampling and 160 testing is limited to ensuring that food sold on the market with an organic claim, is free of 161 pesticide residues. A recent unpublished BSc thesis at the University of Kassel revealed that 162 80% of the samples by CBs in ten EU member countries are taken of final products, but only 163 20% from the field or during the production process. 164 The differentiation between active use and non-intentional contamination is difficult, though, if 165 only final products are tested. Plant (mainly leaf) samples from the field have several ad-166 vantages in this regard: (a) Often, there is a long time span between pesticide application 167 and harvest. Because of dissipation of the residues, nothing or only traces may be found in face/weight ratio between 10 and 118 cm 2 /g 28 , whereas for fruits this ratio is between 0.6 and 171 2.2 29 , and for seeds between 2 and 10 cm 2 /g only 30,31,32 . Residues in leaves are therefore 172 normally higher than in seeds, fruits or roots, which makes interpretation of test results eas-173 ier. (c) Field sampling allows to take separate samples from centre and margin of the field, as 174 explained below in more detail. 175 Unfortunately, if CBs take field samples at all, they often take them only from field mar-176 Other CBs have established so-called "action levels", below which they consider the pres-181 ence of residues in organic products to be the result of ubiquitous environmental contamina-182 tion, with no need to investigate their origin. 33 While such "action levels" may be necessary 183 for specific cases (see below concerning the banana industry), using this approach in a gen-184 eral way disregards not only the spatial distribution, but also temporal dynamics of pesticides 185 in plant tissue. As opposed to soil, half-lives in plant tissue exposed to UV radiation and 186 weather, are relatively short for most modern pesticides 35 . A residue level of 0.02 mg/kg, 187 used by some CBs as "action level", is typically reached one to two months after the applica-188 tion of a pesticide, in some cases even after only five days (Supplementary Table 1). 189

190
The objectives of our study are: (I) to demonstrate that appropriate field sampling methods 191 can differentiate the effects of fraudulent pesticide application by the organic farmer, from the 192 results of both short-range and long-range spray-drift, and (II) for the specific case of aerial 193 fungicide spraying in the banana industry, identify appropriate variables, which allow us to 194 interpret the test results correctly for the purpose of this differentiation. 195 To demonstrate the appropriateness of differentiated field sampling (objective I), in a first part 198 of our study we selected from the CERES archive six cases (Table 1), and one case from the 199 GfRS archive (N° 6 in Table 1). 200 100 to 300 m 2,4-D had been found by a Belgian importer in cocoa beans. Conventional banana farms with a high level of drift in the neighbourhood. 2,4-D is selective for control of dicotyledonous weeds, mainly in cereals, but is frequently used in Latin America also for weed control in perennial crops such as cocoa.

Bulgaria Oil-bearing roses
Leaves 200 to 600 m The inspector had been made believe that a risk of spray-drift did not exist, because conventional neighbour fields were semi-abandoned. Therefore, he had taken only one sample from the centre of the farm, composed of several sub-samples. Later, it turned out that one neighbour had sprayed his rose plantation, using air-blast equipment, four days before the organic field had been sampled.  conditions to discriminate between fraudulent application and spray-drift (objective II), a total 219 of 476 residue tests from Ecuador from 2018 and 2019 were analysed. From these, 24 were 220 excluded because of sampling mistakes. In most cases, to reduce testing costs, first a mixed 221 sample from border and centre is tested ( Supplementary Fig. 3). Based on the assumption 222 that, due to the overall heavy drift problem, residues in mixed samples below 0.1 mg/kg were 223 derived from drift and did not require separate testing, 119 mixed samples were identified as 224 "drift" and no separate testing of margin and centre done. These were excluded from the sta-225 tistical analyse. Residues in 20 mixed samples were so high that they were immediately 226 identified as "application" and were also not considered. This left datasets from 67 farms with 227 222 individual samples (i.e. 67 centre and 155 border samples), which were analysed sepa-228 rately and then subjected to statistical analyses. Of the 67 cases, 14 had been identified as 229 "application", 48 as "drift", while five had remained "unclear". only "Traces < LOQ" for a specific substance, a default value of 0.005 mg/kg was used instead. For testing the variables, multivariate statistical analysis based on logistic regression, 234 discriminant analysis and support vector machines were performed to find rules that would 235 classify farms into the two groups. 37 Here, we only report results of the discriminant analysis, 236 which provided the most satisfactory results. The variable selection and performance of mod-237 els and their prediction accuracy were assessed using leave-one-out and k-fold cross valida-238 tion. The selected variables are graphically displayed using a biplot. 38 A one-way ANOVA 239 was used to test the statistical significance of these variables between the "application" and 240 the "drift" farms. The analysis was performed in R programming language using MASS, caret 241 and klaR libraries. 242

245
The organic apple orchard from Chile has borders with a conventional cherry plantation. 246 While in the sample taken close to the cherries, 11 different pesticides were found, with dif-247 ferent residues adding up to 13.41 mg/kg (the highest value for a single substance was 9.1 248 mg/kg for captan), at 100 m distance only seven substances were found, with a concentra-249 tion of only 3% of the border sampleleaving no doubt that the residues were derived from 250 drift. Yet, the values were so extremely high that the orchard lost its organic status under 251 NOP, while under the Chilean organic standard, the farmer had to establish broad buffer 252 zones. In a nearby blueberry plantation, however, we had the opposite picture: the concen-253 tration of imidacloprid in the margin sample was 0.15 mg/kg, while in the centre of the field, 254 at 100 m from the margin, it was 1.8 mg/kg. This was a clear case of fraud, and the farm lost 255 its certification. 256

257
In the centre sample from farmer 1, 0.023 mg/kg chlorpyrifos was detected, and traces of di-258 chlobenil, but no lambda-cyhalothrin. In the margin sample from SH1, however, no chlorpyri-probably there was an overlap of an application (chlorpyrifos) and drift (for the other two substances). In the case of farmer 2, only traces of deltamethrin were found in the margin sam-262 ple, but no residues in the centre, thus this was a clear "drift" situation. From farmer 3, only 263 one sample was taken, because there was no conventional neighbour. The sample had rela-264 tively high residues of fipronil, clearly showing an application (Supplementary Table 3). 265 These results demonstrate that even for small fields of less than 1 ha, the difference between 266 residues derived from spray-drift and from application by the organic farmer can often be es-267 tablished, especially when neighbours use manual knapsack sprayers. As a result, the 268 group's internal control system excluded several member farmers from the group and had to 269 improve its internal member monitoring. 270

271
The insecticides bifenthrin and chlorpyrifos were detected at levels of 0.005 to 0.013 mg/kg 272 in samples from centre and margin, respectively, of a 4 ha rice field ( Supplementary Fig. 2). 273 The on-site inspection did not reveal any evidence for use of these substances by the or-274 ganic farmer. Short-range drift could be ruled out, because in this case the residues in the 275 centre would be expected to be by a factor 10 lower than in the margin sample. Both insecti-276 cides have a low vapour pressure, therefore long-range drift through evaporation is also ex-277 cluded. The residues could theoretically originate from an application two to three months 278 prior to sampling, but also from long-range drift through dust, or contaminated irrigation wa-279 ter. Many conventional rice farmers in the region use these insecticides. Under the principle 280 of "innocent until proven guilty", the farmer remains certified. 281

282
For finding the origin of 2,4-D residues detected in organic cocoa beans, different samples 283 were taken from the cocoa farm. Weed samples from the centre of the plantation had low 284 levels (0.023 mg/kg) of the herbicide 2,4-D, while weed samples from the field margins and 285 cocoa leaf samples were free of residues. Several fungicides found in the cocoa leaves prob-286 ably came from aerial spraying on nearby banana plantations, but this could not have been while spray-drift from manual knapsack sprayers used in between banana plants, with the nozzle turned downwards, is almost zero. Also, long-range drift could be ruled out, because 290 considering the dense canopy of cocoa trees, this would lead to higher residues in the can-291 opy itself than in the weeds growing beneath (Supplementary Table 4). Dry weeds observed 292 by the inspectors in between the cocoa trees provided further evidence of herbicide applica-293 tion on the organic plantation. Therefore, the certificate was suspended in spite of the low 294 residue level. 295 (1) 303

Oil-bearing Roses from Bulgaria
with: x = distance from the target field and y = deposit at distance x, expressed as a fraction 304 of the deposit on the target field, 305 combined with approximate data concerning the impact of wind speed, 10 CERES concluded 306 that the assumption of these residues being derived from drift, was not plausible (Supple-307 mentary Table 5). Penconazole was also detected in a sample of rinse water from the or-308 ganic farm's sprayer, further supporting the presumption that it was a case of deliberate ap-309 plication. The farm lost its organic status. 310

311
Grape leaves were sampled from eight organic vineyards during a period when conventional 312 farmers were applying fungicides for preventing different fungus diseases. Samples from 313 seven farmers had residues with a maximum of 0.75 mg/kg for folpet and 0.52 mg/kg for sub-blast spraying by neighbours and possibly air swirling caused by thermal lift in the hilly landscape, did not allow for a clear distinction between margin and centre samples. On farm N°8, 317 however, the folpet concentration reached 73 mg/kg, clearly indicating a direct application by 318 the organic farmer. This was confirmed later by a sample taken from sprayer rinsing water. 319 This farmer lost the organic status, while the others remained certified. This decision was 320 correct assuming that under the given weather conditions, all farmers in the region had 321 sprayed more or less at the same time, so that drift effects were not confounded with dissipa-322 tion effects. 323 This was confirmed through collector interviews. Due to the pressure from the CB, the com-339 pany implemented strict measures for preventing delivery of nuts from non-certified areas. As company kept its organic status. 344

345
Sampling banana leaves is a time-consuming effort (Figure 4d). On the first plantation, the 346 same six fungicides were found in the centre and in two border samples. Not only the sum of 347 all pesticide residues was substantially higher in the centre than in the margins, but also the 348 values for most individual substances (Figure 4a). This did not leave any doubt that the resi-349 dues were derived from an application by the organic farmer, whose certificate was then sus-350 pended. On the second plantation, however, only the sample taken close to the conventional 351 banana neighbour had residues, while the samples from the centre and close to a plantain 352 orchard were free of residues. The residues were derived from drift and the farm kept its or-353 ganic status (Figure 4b and Supplementary Fig. 4). 354 In a first approach, the discriminant analysis identified six variables as the most promising 375 ones based on a cross-validated stepwise selection procedure (1subcen, 2subrat2, 3max-376 cen, 4maxrat3, 5sumcen and 6sumrat2, see Figure 5). The one-way ANOVA also indicated 377 that the six selected variables are significantly different between the drift and the application 378 Table 7). The biplot based on the first two principal components using 379 these selected variables explains 60 and 17% of the total variation, respectively (Figure 6b).

389
Farms previously considered as having been subject to drift mostly clustered around zero 390 while application farms scattered on the left side of plot with two exceptions clustering around 391 zero. The farms considered unclear are distributed throughout. The raw data for the six varia-392 bles were visualized using a heatmap (Figure 6a). For this, each variable was standardized 393 to a mean of zero and unit variance. The clustering of farms is visualized using a dendro-394 gram based on the Unweighted Pair Group Method with Arithmetic means (UPGMA). The there is non-negligible heterogeneity within groups, confirming the one-way ANOVA results.
Application farms are clustered in the top rows, showing that two farms that had been consid-398 ered subject to drift grouped clearly with the "applicants", whereas three supposed applicants 399 grouped with the spray drift group. Four of the five unclear cases grouped in the application 400 group or at the edge towards the drift group while one unclear case fell in the drift group. In-401 terestingly, farm 61 that groups in Figure 6b with the applicants, falls into the drift group in Fi-402 gure 6a. Taking a closer look at the initial raw data and sampling record for those farms, 403 which visually in both heatmap and biplot appeared to be misclassified as "drift" (cases 42 404 and 48), revealed that these cases should not have been included in the analysis because 405 the field samples had been taken in a wrong way, and that the cases 5, 8 and 14 should 406 have been classified as "drift". Thus, we subsequently re-classified the latter three cases as 407 "drift" and the former two as "unclear", leaving a training dataset containing the 60 farms for 408 which the class had been assigned as either "drift" or "application". The test dataset contains 409 the five cases originally classified as "unclear" (cases 58, 59, 60, 61, 62) and the two cases 410 subsequently removed from the test set (cases 42 & 48). The linear discriminant function 411 performed best in terms of accuracy with the four variables 2subrat2, 3maxcen, 4maxrat3 412 and 6sumrat2 ( Figure 5 and Figure 7). 413 This linear discriminant function was evaluated by cross validation and found to correctly 414 classify a farm as either "drift" or "application" with an accuracy of 93.3%. The leave-one-out 415 cross validation method was used to evaluate the accuracy of the model. In this method, 416 each sample farm was dropped from the test data and then the class of that farm was pre-417 dicted using the discriminant model. The misclassification rate in this cross validation of a 418 "drift" as an "application" farm was 2.1%. This means that for a farm that is truly a "drift farm" 419 there is an estimated probability of 2.1% that this is erroneously classified as an "application" 420 farm. The misclassification rate of an "application" farm as a "drift" farm was estimated at

433
falsely classified as a "drift" farm. Of course, these estimated error rates are themselves sub-434 ject to estimation error, and it is desirable to accumulate data from more farms to stabilize 435 these estimates, as well as the estimates of the discriminant function. It also needs to be 436 taken into account that, as we have explained here, there was some uncertainty regarding 437 correct group membership for some farms that was only revealed by closer scrutiny of the 438 initial statistical analysis. This may mean that the error rates we obtained in cross-validation 439 of the final analysis presented here are on the optimistic side. The continuation of the present 440 work, and especially the accumulation of data from more farms, will help to avoid such wrong 441 assessments in the future. Three out of the five initially "unclear" cases turned out to belong 442 to the "application" group, two to the "drift" group. 443 The linear discriminant function in our analysis is (See Figure 5 and Supplementary The linear discriminant function is also depicted in Figure 7 for the four selected variables. 452 For each pair of variables, the plot shows the separation of the two groups by two different 453 colours, and the placement of individual samples represents the rate of correct classification. 454 4 Conclusions 455 1. In most cases, comparing pesticide residues in leaf samples from field margins close to a 456 possible source of spray-drift, to samples from the centre of the organic field, allows to 457 distinguish the effects of spray-drift from deliberate pesticide use by the organic farm. 458 The method works even in regions with extremely intensive pesticide use and aerial 459 spraying by conventional neighbours. 460 2. The distinction is also possible when it comes to very small fields, where the distance be-461 tween border and centre is shortprovided that manual knapsack (as is normally the 462 case in smallholder setups) or tractor boom sprayers are used. It becomes difficult to im-463 possible on such small fields, when neighbours use air-blast or aerial spraying. 464 3. When residues below approximately 0.03 mg/kg are found evenly spread over the field, it 465 becomes difficult to distinguish long-range drift (from evaporation or wind erosion) from 466 the results of deliberate use several weeks before sampling. In such cases, the test re-467 sults alone do not allow to prove fraudulent practices, as long as other evidence (pesti-468 cide containers in the farm house, residues in rinse water in the sprayer, records, etc.) do 469 not exist. method in a meaningful way. Sampling must be planned in a way that allows for clear interpretation of results. Taking only one sample from a field, often leads to useless results. 473 Sampling residues in spraying equipment, cross-checking with book-keeping records and 474 other inspection methods, should be used as complementary methods. 475 5. A weak point of our survey of the 67 banana farms is that some of the centre samples 476 were not taken at sufficient distance from the edges. This has meanwhile been corrected 477 through improved work instructions ( Supplementary Fig. 5). Another correction of the pro-478 cedure, which is currently being tested, is reduction of the "action level" for separate test-  6. When a reference sample from the field margin is not available, and residues are high in 483 the central part of the organic farm, comparing the test results to expected values from 484 standard deposition curves, can be enough to distinguish drift from application. 485 7. For the specific conditions of aerial fungicide spraying in the banana industry, the varia-486 bles explained in Figure 5 and a linear discriminant function such as the one outlined 487 above can be used tentatively for differentiating drift from application. We suspect that 488 the same method can be used for other crops exposed to heavy drift pressure (e.g., fruit 489 orchards and vineyards), but this is yet to be confirmed.