Comanaging fresh produce for nature conservation and food safety

Pathogens in Fresh Produce.

Since October 2, 2006, an organic farming operation with numerous fields spread across California’s Central Coast region has consistently tested all of the leafy green vegetables that it produced or acquired for foodborne pathogens. At harvest, vegetables were packed into boxes and loaded into pallets, with each 40 × 48 × 78 in pallet holding ∼600 lb of product. The pallets were transported to centralized packing centers and sorted into production units (four pallets per production unit). Sterilized forceps were used to grab 60 “pinches” of fresh leafy greens (∼150 g) from each production unit. Samples were placed into sterile plastic bags and transported to an in-house molecular laboratory operated by a third-party company. Unique farm names were associated with each vegetable sample, and records of spatial farm locations were obtained from food-safety staff at the larger farming operation.

Although established methods for detecting EHEC and Salmonella are readily available (11, 12, 21), the specific methods used by the third-party company were not disclosed. Generally, however, the company’s pathogen testing procedure consists of DNA preparation, amplification, and detection. First, samples were enriched, and bacterial DNA was extracted with a lysing procedure. Then samples were subjected to PCR, amplifying unique sequences from E. coli O157:H7, shiga toxin-producing E. coli (STEC), and S. enterica. PCR products were visualized on agarose gels with ethidium bromide, and positive bands were subjected to molecular confirmation, using four to six more multiplex reactions with and without magnetic bead separation, depending on the pathogen. Any sample that tested positive for E. coli O157:H7 or STEC was labeled as containing EHEC.

Between 2007 and 2013, a total of 482,208 samples originating from 295 farms were tested for E. coli O157:H7, STEC, and Salmonella. The farms were located in the United States (Arizona, California, Colorado, Nevada, and Oregon), Mexico (Baja California, Guanajato, and Sonora), and Chile. Samples were obtained from a variety of leafy green vegetables and cultivars (SI Appendix, Table S4).

Generic E. coli in Water.

Irrigation water was sampled monthly by the leafy green agricultural industry across the California Central Coast region over a 4-y period (2007–2010). A commercial laboratory provided the blinded data, but sample collection locations were preserved using GPS coordinates. Samples were collected from on-farm wells and reservoirs and municipal water distribution systems by either contracted professionals or farm personnel at the points of distribution and application. Samples were captured in 250-mL Whirl-Pak bags and transported on ice (∼4 °C) for microbial analysis.

The samples were analyzed for the presence of generic E. coli using a most probable number (MPN) technique; specifically, an enzyme substrate test typically performed using a commercially prepared quantification tray. In this test, 100 mL of sample water was added to a single tray and incubated at 35 ± 0.5 °C for 24–28 h. The presence of E. coli was determined by the presence of an enzymatic reaction that caused a color change in the substrate. Results were entered into an MPN calculator for final quantification, with lower and upper bounds of detection from 1 to 2,419.6 MPN/100 mL.

Salmonella in Rodents.

Fecal specimens were collected from 11 wild rodent species trapped in nine produce farms located in Monterey and San Benito Counties, California, between October 2009 and August 2011, as described previously (21). Trapping was carried out approximately once every 3 mo at each farm over the study period. Smaller-sized nocturnal rodents (e.g., Peromyscus maniculatus, Peromyscus californicus) were trapped using Sherman live traps, and California ground squirrels (Spermophilus beecheyi) were trapped using Tomahawk live traps. Fresh fecal pellets and bedding material were transported on ice to the University of California Davis for pathogen detection. All animals were handled humanely in accordance with the Institutional Animal Care and Use Committee’s protocol no. 16376.

To detect Salmonella, an aliquot of each fecal sample (∼0.025–0.10 g per rodent) was preenriched in 50 mL of buffered peptone water (BD Diagnostic Systems) for 20 h at 37 °C. Then 10 mL of preenrichment broth was transferred to 1 mL of Rappaport Vassiliadis medium (BD Diagnostic Systems) for 48 h at 42 °C. A loopful of broth was streaked for isolation onto xylose lysine desoxycholate agar (BD Diagnostic Systems). A pure isolated colony was confirmed biochemically using lysine, citrate, triple sugar iron, and urea (all from BD Diagnostic Systems).

Pathogen Prevalence over Time.

We assessed whether the prevalence of EHEC and Salmonella changed over time with the leafy greens dataset. First, we calculated the number of detections and absences of each pathogen in each year. Then we modeled how the proportion of positive pathogen detections changed from 2007 to 2013, using linear mixed models with survey years as random effects (37). Proportions of EHEC and Salmonella detected were fourth root-transformed and square root-transformed, respectively, to satisfy model assumptions (e.g., normality, heteroscedasticity). To determine whether yearly dynamics differed among the California Central Coast, other counties in California, and other surveyed states/countries (see above), we included the following as fixed effects: ordinal survey year, an indicator variable demarcating the location of the pathogen test, and their interaction. We evaluated significance through backward model selection, comparing nested models with likelihood ratio tests (37).

We also assessed seasonal variation in pathogen prevalence at farms in the Central Coast region. First, we obtained the Julian day for which each sample was harvested. Then, for each day of the year, we calculated the fraction of samples that tested positive for EHEC and Salmonella (across all sites and years). Days with fewer than 100 pathogen tests were excluded. We then constructed linear models in which Julian date and Julian date² were included as fixed effects. As before, we assessed significance with backwards model selection.

Creating Land Cover Maps.

To determine how landscape features influence pathogen prevalence, we created an Anderson level II terrestrial land use/land cover map of California’s Central Coast region. Vegetation in 2005 and 2012 was mapped within a 1.5-km buffer of all agricultural areas in Monterey County’s Salinas Valley (1,669.13 km²). The maps also encompassed parts of the San Juan and Pajaro Valleys in Santa Cruz and San Benito Counties that were strategically chosen to encompass sites sampled for pathogens (2005: 81.33 km²; 2012: 236.97 km²). Vegetation was hand-classified into one of 16 possible categories from 1-m² resolution National Agricultural Inventory Program imagery, obtained in the summers of 2005 and 2012 (SI Appendix, Table S5).

All cropland and riparian habitat was mapped with a 0.5-acre minimum mapping unit (MMU). Upland vegetation, urban areas, and other land use classes were mapped with a 1.0-acre MMU. To assess the effects of proximity to grazeable land, we overlaid spatial grazeable land data from 2010 on top of our land cover map, which was acquired from the California Department of Conservation (www.conservation.ca.gov/dlrp/fmmp/Pages/Index.aspx). Grazeable land was defined as “land on which the existing vegetation is suited to the grazing of livestock” and identified in collaboration with several organizations, including the California Cattlemen’s Association and the University of California Cooperative Extension.

To determine how food safety concerns may have affected land cover, we compared the 2005 (preoutbreak) and 2012 (postoutbreak) land cover maps. We isolated agriculture-induced change rather than natural processes (such as riparian vegetation regeneration) by narrowing our focus to only land within 50 m of 2005 cropland (5). We quantified the changes in bare ground, cropland, riparian areas, and other natural vegetation across the Salinas Valley between 2005 and 2012 (SI Appendix, Table S5).

Landscape Associations with Pathogens.

To model the effects of different land covers on pathogen prevalence, we calculated the fractions of cropland, grazeable land, riparian areas, and other natural vegetation classes (i.e., woodland, grassland, upland scrub, and meadow/marsh) around each pathogen survey site at multiple scales (within 500, 750, 1,000, 1,250, and 1,500 m of survey sites). We then assessed the effects of cropland, grazeable land, riparian areas, and other natural vegetation at each spatial scale with separate linear models. Recent studies have indicated that generalized linear mixed models (with Poisson, negative binomial, or binomial error structures) can have high type I error rates (38). Indeed, our preliminary analyses using generalized linear mixed models with binomial errors yielded type I error rates of >60%. Therefore, for all analyses, we used linear mixed models to quantify changes in the proportion of samples that were positive for each pathogen.

For analysis of generic E. coli in water, we first converted from an MPN to presence/absence (using a MPN of 1 as a threshold). We then calculated the fraction of samples that tested positive for generic E. coli across all years at each farm, square root-transforming proportions to satisfy model assumptions. Farms without at least 50 E. coli tests were excluded from the analyses, but our results were robust to varying this threshold. We next averaged the land cover values for all survey points within a given farm so that each farm was assigned one average value for each land cover type. Cropland, grazeable land, riparian areas, and other natural vegetation were included as fixed effects in each model, and closest city was included as a random effect to account for spatial autocorrelation.

For analyses of both EHEC and Salmonella in leafy greens, we first calculated the fraction of samples testing positive across all years at each farm, and then excluded farms without at least 100 pathogen tests. As before, our results were robust to varying this threshold. Proportion data were either fourth root-transformed (EHEC) or square root-transformed (Salmonella), so that resulting models would yield normally distributed residuals. For both pathogens, the four land cover types were included as fixed effects in all models, and “closest town” nested within “county” were included as random effects.

Finally, owing to fewer sampling locations, we quantified the fraction of rodent samples that tested positive for Salmonella in each year at each farm. Site-years without at least 10 samples were excluded from the analysis. As for the water dataset, we calculated average land cover values across all survey locations within each farm. The four land cover types were included as fixed effects in all models, and survey year was included as a random effect.

For each analysis, we used model averaging to assess the association of each land cover type (cropland, grazeable land, riparian vegetation, and other natural vegetation) with pathogen prevalence and reported nonshrinkage variance estimates. Nonshrinkage variance estimates for a fixed effect are estimated from a combination of model uncertainty and standard regression variance, using only the models that include that given fixed effect (39). All analyses were performed in R (40), using “lme4” for linear mixed models and “MuMIN” for model averaging.

Detecting Effects of Land Use Change.

We assessed whether temporal changes in vegetation management practices were associated with temporal changes in EHEC and Salmonella prevalence in leafy green vegetables on the 28 farms that were regularly surveyed for pathogens and had their surrounding land cover mapped in 2005 and 2012. First, we calculated the change in the fraction of samples that tested positive for EHEC and Salmonella between early surveys (2007–2008) and late surveys (2011–2013). Next, for each farm, we calculated the change in cropland, riparian areas, and other natural vegetation between 2005 and 2012 at each spatial scale. Given its lower spatial resolution and temporal mismatch with the land cover map, we did not assess changes in grazeable land; instead, we examined the effects of changes in the extent of cropland, riparian, and nonriparian vegetation regardless of whether it was grazed or ungrazed.

Finally, we modeled associations between land cover changes and pathogen prevalence changes between the early and late survey periods using linear mixed models, acknowledging that other factors may contribute to some of the temporal shifts in pathogen prevalence (e.g., adoption of industry-wide reforms; ref. 4). Because seminatural vegetation was often replaced with crops, land cover variables were highly collinear; thus, we constructed separate models to investigate each land cover type, pathogen (EHEC and Salmonella), and spatial scale (500, 750, 1,000, 1,250, and 1,500 m). “Nearest town” nested within “county” were included in all models as random effects to account for spatial autocorrelation. For models predicting associations between EHEC and cropland or natural vegetation, we also included an exponential variance structure to account for heteroscedastic residuals (37). Significance was assessed for each model with backward model selection (37).