Landscape and Meteorological Factors Affecting Prevalence of Three Food-Borne Pathogens in Fruit and Vegetable Farms

Abstract

Produce-related outbreaks have been traced back to the preharvest environment. A longitudinal study was conducted on five farms in New York State to characterize the prevalence, persistence, and diversity of food-borne pathogens in fresh produce fields and to determine landscape and meteorological factors that predict their presence. Produce fields were sampled four times per year for 2 years. A total of 588 samples were analyzed for Listeria monocytogenes, Salmonella, and Shiga toxin-producing Escherichia coli (STEC). The prevalence measures of L. monocytogenes, Salmonella, and STEC were 15.0, 4.6, and 2.7%, respectively. L. monocytogenes and Salmonella were detected more frequently in water samples, while STEC was detected with equal frequency across all sample types (soil, water, feces, and drag swabs). L. monocytogenes sigB gene allelic types 57, 58, and 61 and Salmonella enterica serovar Cerro were repeatedly isolated from water samples. Soil available water storage (AWS), temperature, and proximity to three land cover classes (water, roads and urban development, and pasture/hay grass) influenced the likelihood of detecting L. monocytogenes. Drainage class, AWS, and precipitation were identified as important factors in Salmonella detection. This information was used in a geographic information system framework to hypothesize locations of environmental reservoirs where the prevalence of food-borne pathogens may be elevated. The map indicated that not all croplands are equally likely to contain environmental reservoirs of L. monocytogenes. These findings advance recommendations to minimize the risk of preharvest contamination by enhancing models of the environmental constraints on the survival and persistence of food-borne pathogens in fields.

INTRODUCTION

Despite the implementation of produce safety practices, food-borne outbreaks associated with fresh produce continue to result in significant illnesses, hospitalizations, and deaths. Approximately 13% of reported food-borne outbreaks were linked to produce from 1990 to 2005 (1). While this increase in produce-associated outbreaks may, in part, be due to improved surveillance of produce commodities, fresh produce will remain a vehicle for food-borne disease for two reasons: (i) the consumption of fresh produce commodities has increased due to promotion of a healthy lifestyle associated with eating fresh produce, and (ii) fresh produce commodities are often consumed raw. As a result, contamination with food-borne pathogens at any point in the supply chain from farm to fork has a heightened chance of causing disease (2). Three major bacterial food-borne pathogens, Listeria monocytogenes, Shiga toxin-producing Escherichia coli (STEC), and Salmonella, have been associated with disease outbreaks linked to produce. Together, these pathogens account for an estimated 76% (653/861) of deaths attributed to known bacterial food-borne pathogens in the United States (3).

STEC and Salmonella have been responsible for the majority of produce-associated outbreaks (1, 4, 5). Many of these outbreaks were traced back to environmental reservoirs located on the implicated farms. In 2006, one of the first major produce-associated outbreaks linked to preharvest contamination was in spinach. The outbreak-associated strain of E. coli O157:H7 was isolated from feral swine, cattle, surface water, sediment, and soil (6). An investigation of an outbreak of Salmonella enterica serovar Saintpaul in jalapeño peppers recovered the outbreak strain from peppers in the field and irrigation water from one of the implicated farms (7). L. monocytogenes has the potential to cause produce-associated outbreaks (8, 9). In 2011, a cantaloupe-borne L. monocytogenes outbreak caused 146 illnesses, 30 deaths, and 1 miscarriage in 28 U.S. states (10). The source of the outbreak was suspected to be a piece of contaminated equipment in the farm packinghouse (11). Although the cause of this outbreak was not due to preharvest contamination of the melons, the results of the outbreak investigation demonstrate the potential risk of L. monocytogenes contamination in produce and the difficulties associated with managing this pathogen in the food safety system, because incidental contamination originating from food or environmental sources can persist through food-processing/handling facilities (12).

Studies have demonstrated the ability of food-borne pathogens to survive for extended periods of time in the soil and water, with the potential to infect new hosts and/or contaminate food products (13–15). Laboratory and field studies have identified a number of likely sources of preharvest contamination, such as irrigation water, application of untreated manure, runoff water from livestock operations, and wildlife intrusion into fields (16, 17). Management of farms at the farm or production block scale might greatly influence the local movements of the pathogens and the chance for produce contamination. However, contamination of produce in the preharvest environment remains a complex challenge because the conditions that promote persistence of pathogens in the preharvest environment and subsequent produce contamination are not well understood. Each farm landscape is a unique combination of numerous environmental characteristics that we hypothesize set the baseline conditions for persistence of pathogens in or near produce fields.

The focus of the presented research was to better understand L. monocytogenes, STEC, and Salmonella in the produce preharvest environment and, more specifically, to identify specific, remotely sensed, topographical properties (e.g., proximity to forests), soil properties (e.g., available water storage), and meteorological events (e.g., precipitation) that may influence pathogen prevalence. To that end, we assessed the prevalence, persistence, and diversity of food-borne pathogens among farms, seasons, and sample types. Classification tree (CT) models were used to identify remotely sensed landscape (i.e., topographic and edaphic) and meteorological characteristics that delineate the presence and absence of food-borne pathogens in the preharvest environment (18, 19). By modeling food-borne pathogen contamination as an ecological process, we seek to supply food safety researchers and professionals with recommendations to minimize the risk of preharvest contamination.

MATERIALS AND METHODS

Field sampling design.

A longitudinal field study was performed on five produce farms in New York State (NYS). Farms were selected on the basis of the willingness of growers to participate and to sample farms geographically distributed across NYS. Farms were not selected on the basis of management practices. Farms were sampled nine times from June 2009 to August 2011. Farms were located in three regions of NYS: central New York (n = 1), the Finger Lakes (n = 3), and western New York (n = 1). The distance between farms ranged from 33 to 205 km. Sample size calculations were performed using the lower end of the reported range of prevalence estimates for L. monocytogenes (20), Salmonella (21), and STEC (6) in order to reach 50 isolates for each targeted pathogen. However, due to time and budget constraints, 588 samples that yielded 107, 27, and 16 representative L. monocytogenes, Salmonella, and STEC isolates, respectively, were collected. Farms were sampled every astronomical season (summer, fall, winter, and spring). Samples were not collected during snow cover in winter.

Within each farm, four fields were selected to standardize sample sizes among farms since the overall farm sizes varied considerably. Fields that had produce commodities generally consumed raw were selected and were selected to capture topographical field diversity, such as low and high elevations in the field. During each sampling excursion, a single soil sample consisting of five subsamples of topsoil from five locations in each field was collected. Soil samples were pooled because pathogens were expected to have high spatial variability and small population sizes within the fields (22, 23). One area drag swab sample and, where available, up to five water and fecal samples were collected for each field. In total, 77 (68 surface and 9 engineered), 9 (all engineered), 45 (44 surface and 1 engineered), 18 (9 surface and 9 engineered), and 25 (all surface) water samples were collected from each of the five farms. Fecal samples represented only 10% of the 588 total samples, and the majority of fields did not contain feces. Global positioning system (GPS) coordinates were recorded for each sample collected within the field and revisited upon each subsequent visit, in order to access possible persistence of the targeted food-borne pathogens in the preharvest environment. General farm characteristics were documented (Table 1).

Table 1.

General farm and key management characteristics

Farma	Size (acres)	Organic	Irrigateb	Manurec	Compost or composted manured	Staff (no. of employees)
						Year-round	Temporary
1	>1,000	No	No	Yes	Yes	Yes (31–40)	Yes (40+)
2	<250	Yes	Yes	Yes	Yes	No	Yes (1–5)
3	>1,000	No	Yes	Yes	No	Yes (6–10)	Yes (21–30)
4	<250	No	Yes	No	No	No	Yes (11–15)
5	>1,000	No	No	No	No	Yes (11–15)	Yes (31–40)

^aAll farms answered “yes” to wildlife control measures (e.g., hunting or fences), worker training (e.g., sessions or videos on good hygiene and sanitation practices), and having good agricultural practices (GAP) plans (i.e., third-party audits of food safety practices).

^bFarm 2 used a combination of drip and overhead irrigation, depending on the crop, farm 3 used overhead irrigation, and farm 4 used drip irrigation.

^cManure slurry had been applied to a field within the previous year. Produce was not planted before 120 days.

^dCompost and composted manure are treated products.

Sample collection.

Latex gloves and disposable plastic boot covers (Nasco, Fort Atkinson, WI) were worn for sample collection. Gloves and boot covers were changed between each field, and gloves were disinfected with 70% ethanol prior to sample collection. A total of 588 samples were collected. Approximately 6-inch (15.2-cm)-deep soil samples and fecal deposits were gathered into sterile Whirl-Pak bags (Nasco, Fort Atkinson, WI) using sterile scoops (Fisher Scientific, Hampton, NH). Premoistened drag swabs, as described by Uesugi et al. (14), were dragged around the field perimeter and diagonally back and forth for ≥10 min, covering an average field area of 0.75 ha. Drag swabs were deposited into a Whirl-Pak sample bag containing 45 ml of phosphate-buffered tryptic soy broth (pTSB; Becton, Dickinson, Franklin Lakes, NJ). Water samples were collected directly into sterile 250-ml jars by use of a 3.66-m sampling pole (Nasco, Fort Atkinson, WI). These water samples were taken a minimum of 2 m from the shoreline and 0.3 m below the surface. All samples were transported on ice, stored at 4 ± 2°C, and processed within 24 h of collection.

Sample preparation.

All samples were used for three separate enrichment schemes to allow the isolation and detection of (i) L. monocytogenes, (ii) E. coli O157:H7, and (iii) a combined enrichment for non-O157 STEC and Salmonella. The five soil samples collected in each field were weighed into 5-g portions, and the portions were combined to form a 25-g pooled sample. Three 25-g pooled soil samples were prepared and deposited in sterile filter Whirl-Pak bags. For fecal samples, 2 to 10 g of each fecal sample collected was deposited into three sterile filter Whirl-Pak bags. Drag swab samples were mixed with 90 ml pTSB in the Whirl-Pak bag by hand massaging for 2 min. The drag swab was then squeezed, and 10 ml of the liquid contents from the sample bag was aseptically transferred to each of three sterile filter Whirl-Pak bags. Water samples were processed according to Environmental Protection Agency (EPA) standard methods (24, 25). Each water sample collected (250 ml) was passed through a 0.45-μm-pore-size filter unit (Nalgene, Rochester, NY). This filter was then aseptically removed and cut into three equal-sized pieces, and the pieces were transferred to separate sterile filter Whirl-Pak bags.

L. monocytogenes enrichment and isolation.

L. monocytogenes detection and isolation from the environmental samples collected were performed as in previous studies (26–28). Briefly, samples were diluted 1:10 with buffered Listeria enrichment broth (BLEB; Becton, Dickinson, Franklin Lakes, NJ). These enrichments were incubated at 30 ± 2°C for 4 h. At 4 h, Listeria selective enrichment supplement (Oxoid, Cambridge, United Kingdom) was added. After 24-h and 48-h incubations at 30 ± 2°C, 50 μl of each enrichment was streaked onto modified Oxford agar (MOX; Becton, Dickinson, Franklin Lakes, NJ) and L. monocytogenes plating medium (LMPM; Biosynth International, Itasca, IL). MOX and LMPM plates were incubated for 48 h at 30 and 35 ± 2°C, respectively. Up to 10 L. monocytogenes presumptive colonies were substreaked to brain heart infusion (BHI) agar (Becton, Dickinson, Franklin Lakes, NJ). BHI agar plates were incubated for 37 ± 2°C for 24 h. Presumptive L. monocytogenes colonies were confirmed by PCR and partial sigB gene sequencing (29–31).

E. coli O157:H7 enrichment and isolation.

Samples in Whirl-Pak bags were diluted 1:10 with pTSB and incubated for 2 h at room temperature (23 ± 2°C) to aid in the recovery of injured cells (32). Enrichments were transferred to 42 ± 2°C and incubated for 24 h. Enrichments were subjected to immunomagnetic separation (IMS) to concentrate E. coli O157:H7 cells as previously described (33). Washed IMS beads (50 μl) were plated onto two selective and differential media: modified sorbitol-MacConkey agar (mSMAC; Becton, Dickinson, Franklin Lakes, NJ) supplemented with 20 mg/liter of novobiocin and 2.5 mg/liter of potassium tellurite (Sigma-Aldrich, St. Louis, MO) and CHROMagar O157 agar (CHROMagar, Paris, France). CHROMagar O157 and mSMAC plates were incubated at 37 ± 2°C for 24 and 48 h, respectively. Up to 10 presumptive E. coli O157:H7 colonies were substreaked onto BHI and incubated at 37 ± 2°C for 24 h. Presumptive E. coli O157:H7 colonies were confirmed using a multiplex PCR assay that simultaneously screens for hlyE, fliC_H7, eaeA, rfbE, stx₁, and stx₂, as previously described (34, 35).

Non-O157 STEC and Salmonella enrichment and isolation.

The nonselective enrichment step (pTSB) for both non-O157 STEC and Salmonella was the same. Environmental samples were diluted 1:10 with pTSB and incubated for 2 h at 23 ± 2°C, followed by 24 h of incubation at 35 ± 2°C.

To isolate non-O157 STEC, a 1-ml aliquot of the nonselective enrichment was transferred to 9 ml of EC broth (Oxoid) and incubated at 37°C with shaking for 24 h. A 50-μl aliquot of E. coli broth was plated onto washed sheep's blood agar (Hemostat, Dixon, CA) with 10 mM CaCl₂ and 0.5 μg/ml mitomycin C (WBMA; Fisher Scientific, Hampton, NH) and incubated at 35 ± 2°C for 24 h. Up to 20 colonies that demonstrated enterohemolysis were substreaked to sorbitol MacConkey agar (SMAC) plates and incubated at 37 ± 2°C for 24 h. Up to 10 colonies that rapidly fermented sorbitol were substreaked to BHI and incubated at 37 ± 2°C for 24 h. Presumptive non-O157 STEC colonies were confirmed by the multiplex PCR described above (35) and considered positive if one or both stx genes were detected.

Salmonella detection and isolation were performed using a modified version of the procedures outlined in the Food and Drug Administration's Bacteriological Analytical Manual (36). Aliquots of nonselective pTSB enrichment of 1.0 and 0.1 ml were transferred to 9 and 9.9 ml of tetrathionate (TT; Oxoid) and Rappaport Vassiliadis medium (RV; Oxoid, Fisher, Acros Organic, Belgium), respectively. These selective enrichment cultures were incubated in a shaking water bath at 42 ± 2°C for 24 h. A 50-μl aliquot of selective enrichment was plated onto xylose lysine deoxycholate agar (XLD; Neogen, Lansing, MI) and CHROMagar Salmonella agar (CHROMagar) and incubated at 35 and 37 ± 2°C, respectively, for 24 and 48 h, respectively. Up to 20 presumptive Salmonella colonies were substreaked to BHI and incubated at 37 ± 2°C for 24 h. Presumptive Salmonella colonies were confirmed using a previously described PCR assay that detects invA, a gene specific to Salmonella enterica (37).

Controls and storage.

Positive and negative controls were processed in parallel with each pathogen detection and isolation scheme. The following strains were used as positive controls: FSL R3-001 for Listeria monocytogenes (actA deletion mutant [38]), strain ATCC 43895 tagged with green fluorescent protein (GFP; FSL F6-825) for E. coli O157:H7 (39), FSL F6-704 for non-O157 STEC (E. coli O26:H11), and strain ATCC 700408 tagged with GFP (FSL F6-826) for Salmonella (56) Negative controls were sterile enrichment media. All isolates were preserved at −80°C in 15% glycerol.

Characterization of isolates.

All L. monocytogenes, STEC (E. coli O157:H7 and non-O157 STEC), and Salmonella isolates were streaked from frozen culture onto BHI and incubated at 37°C for 18 h, and a well-isolated colony was selected. Nucleotide sequences of sigB from L. monocytogenes isolates were obtained by Sanger sequencing, preformed by the Cornell University Life Sciences Core Laboratories Center, and compared with those in the GenBank database using BLASTN to assign allelic types, i.e., a unique combination of polymorphisms (40, 41). L. monocytogenes isolates that shared the same allelic type from the same location at least three times were considered possible persistent subtypes and were further subtyped by pulsed-field gel electrophoresis (PFGE). PFGE typing was performed using the standard CDC PulseNet protocol with the restriction enzymes AscI and ApaI (42). Salmonella serovar Braenderup digested with XbaI was used as the reference standard, which allowed normalization and comparison of gel images (43). Pattern images were captured with a Bio-Rad Gel Doc system and the Multi-Analyst software (Bio-Rad Laboratories, Hercules, CA). PFGE banding patterns were analyzed using BioNumerics software (Applied Maths, Saint-Matins-Latem, Belgium). Comparisons were performed using similarity analyses by using the unweighted pair group-matching algorithm (UPGMA) and the Dice correlation coefficient with a maximum space tolerance of 1.5%.

To confirm the identity of O157:H7 isolates and to determine the serotype of non-O157:H7 STEC isolates, comprehensive O serotyping and H typing were performed on one representative STEC isolate per positive sample at the E. coli Reference Center at Pennsylvania State University (State College, PA), as previously described (44, 45).

Salmonella cultivation methods use four combinations of selective enrichments and plating media. To account for the possibility of different strains of Salmonella being isolated from the sample, one representative isolate per Salmonella-positive sample from each isolation scheme (e.g., TT to XLD or RV to XLD) was selected for molecular subtyping (46). Serotyping and PFGE were performed on isolates selected. Serotyping, using the White-Kauffman-Le Minor scheme (formerly known as the Kauffman-White scheme), was performed by the Wadsworth Center, New York State Department of Health (Albany, NY) (47). PFGE typing was performed according to the standard CDC PulseNet protocol for Salmonella using the restriction enzyme XbaI (48).

Descriptive data analysis.

Associations of pathogen-positive cultures with farm, season, or sample type were determined using a chi-square test. A Fisher exact test was used if the expected frequency in any cell was less than 5. Confidence intervals were calculated assuming a binominal distribution. Individual P values were calculated and were considered statistically significant if less than 0.05. Bonferroni's correction was used to account for multiple testing of the three statistical hypotheses (farm, season, and sample type) (49). The diversity of subtypes within farm, season, and sample type was quantified using Simpson's index of diversity (D) (50). All statistics for descriptive analyses were performed in SAS (version 9.1) software (SAS Institute Inc., Cary, NC).

Topographical and soil (spatial) data.

Spatially dependent predictor data were obtained for each sample site (see Table S1 in the supplemental material). GPS coordinates of samples were imported into the Geographical Resources Analysis Support System (GRASS) geographic information system (GIS) environment (51). Site coordinates were projected from latitude-longitude into the Universal Transverse Mercator (UTM) coordinate system, North American Datum of 1983. Map layers for land cover (National Land Cover Database [NLCD], 2006) and the digital elevation model (DEM; Shuttle Radar Topography Mission, 1-arc-second data set) were acquired from the U.S. Geological Survey (USGS) EarthExplorer geographical data bank (http://earthexplorer.usgs.gov/). Map layers for soil characteristics were acquired from the U.S. Department of Agriculture Soil Survey Geographic (SSURGO) database (http://soils.usda.gov/survey/geography/ssurgo/). Road and hydrologic line graphs were obtained from the Cornell University Geospatial Information Repository (CUGIR; http://cugir.mannlib.cornell.edu/). Proximity data were derived from the NLCD land cover base map by calculating the Euclidean nearest-neighbor distance to the desired land cover type. Proximity to urban areas was calculated from a map combining road lines with all four classes of developed land cover. Proximity to water was calculated from a map combining water body areas and flow lines. Percent slope was derived from the DEM. In total, 15 different landscape factors were obtained for CT model development, such as soil type; slope; drainage class; available water storage; organic matter; and proximity to urban development, pastures, forests, and water (see Table S1 in the supplemental material).

Meteorological (temporal) data.

For each sample collection date, meteorological variables were obtained from the major airport nearest each farm, using the airport weather stations in the National Oceanic and Atmospheric Administration (NOAA) National Climate Data Center (NCDC) Local Climatology Database (http://gis.ncdc.noaa.gov/map/lcd/). A major airport was within 60 miles of each farm used in the study. While small-scale differences in weather may be observed between airport and farm, the study aimed to capture the association between remotely sensed meteorological data and pathogen prevalence. In total, 70 different meteorological factors were obtained for CT model development, including temperature (maximum, minimum, and daily average) and precipitation amounts (see Table S1 in the supplemental material). Direct measures of temperature and precipitation were acquired for the day of sampling and 3 days antecedent. The average temperature and precipitation were calculated for each time period, ranging from 1 to 10 days prior to sample collection. Frost cycles were counted by summing the number of times that the surface air temperature fluctuated above and below 0°C for each time period ranging from 1 to 10 days prior to sampling. Averaging and frost calculations were performed using a Perl script (code available from Peter W. Bergholz).

Spatial and temporal data analysis.

The methods used in our analysis of spatial and temporal factors were adapted from Ivanek et al. (19). Large numbers of landscape and meteorological variables were included in our classification analysis as possible predictors of pathogen presence (see Table S1 in the supplemental material). Since there was high potential for covariation among landscape and meteorological predictors of pathogen presence, detrending and principal components analysis (PCA) techniques were applied to account for the linear covariation among predictors. PCA was performed using the ade4 package in the R (version 2.13.1) program (52).

It was desirable to account for season, temperature, and precipitation as independent factors predicting pathogen presence, but season and the meteorological variables did not behave independently at monthly time scales. Temperature and precipitation were detrended for the seasonal effect by performing linear regressions and retaining the residuals from these regressions to represent variation of temperature and precipitation within seasons. Soil properties and elevation were also dependent on the general farm properties, so to examine the effects of soil property and elevation variation within farms, these characteristics were detrended against farm using linear regressions.

Detrended residuals were standardized and used as input for two PCAs to synthesize variation among meteorological and landscape data, respectively, into eigenvectors representing the characteristic behavior of these variables. PCA on meteorological variables yielded an eigenvector that represented 56.1% of the total variation and corresponded well to all temperature variables except average temperature 3 days prior to sampling. The same PCA yielded a second eigenvector describing 18.4% of the total variation that corresponded well to all precipitation variables except precipitation on the day before and the day of sampling. A second PCA showed that landscape data demonstrated less covariation among landscape data. This PCA yielded a single useful eigenvector, representing 51% of the total variation. Available water storage and soil organic matter properties were loaded on this eigenvector, but topographic data were retained as independent predictors of pathogen presence. These three eigenvectors were used as predictor variables in CT models, as they synthesized characteristics of multiple, covarying temperature, precipitation, and soil variables, respectively. This allowed us to minimize the number of predictor variables in the CT models to those that behaved independently.

CT model development.

Tree-based modeling was used to determine rules that classified sampled sites by pathogen presence or absence. Splits were formed by maximizing the homogeneity of presence-versus-absence results in each node according to the Gini index (18). The CTs were built using the rpart package in R (version 2.13.1) (53). To assess the predictive power of resulting trees, a cross-validation procedure was performed 25 times for each tree. The detection methods for food-borne pathogens in the environment are not 100% sensitive or specific; therefore, the response variable was weighted to maximize the predictive power of the resulting tree. To limit the potential effect of different CT outcomes on the basis of weighting of the response variable, we performed a sensitivity analysis in which different weights were applied to negative samples to reflect probabilities of false negatives. The weight of positive samples was always set to 1. The weights for negative results were varied until the weight that minimized cross-validation error was discovered. This weighting scheme was used to produce CTs with 25-fold cross-validation. CTs were pruned to the number of splits that minimized cross-validation error within the selected weighting scheme. This combination of procedures resulted in CTs of reproducible size, predictive power, and split rules, and the subsequent analysis of L. monocytogenes results by random forest using the CT model supported the CT outcome entirely (data not shown) (54).

Geospatial search for L. monocytogenes reservoirs.

Classification trees and related techniques result in rules that can be used to predict the most likely areas to observe a species (18). Using L. monocytogenes as an example, the rules from the L. monocytogenes CT were applied in a GIS framework to explore the potential for croplands to harbor persistent L. monocytogenes in a central New York State landscape (see Results for rule definitions). All calculations on maps were performed using the GRASS GIS (version 6.4.1). Raster maps of (i) water features and flow lines and (ii) pasture areas were extended to reflect proximity-based split rules from the CT using the spatial buffering function r.buffer. These rasters were then converted to vector maps and used in vector map queries using v.overlay to determine cropland areas corresponding to three categories of hypothetical L. monocytogenes reservoir on the basis of CT results. Reservoir polygon areas and minimum reservoir distance from pasture-class land areas were calculated using statistical functions in the GRASS GIS.

Accession numbers.

Isolate information and subtyping data from this study are archived and available through the Food Microbe Tracker database using a guest user login. To facilitate batch retrieval of isolate records, the accession numbers have been tabulated (see Table S4 in the supplemental material).

RESULTS

L. monocytogenes prevalence.

L. monocytogenes prevalence was estimated to be 15.0% (88/588) across all samples collected. Farm, season, and sample type were found to be significantly associated with the frequency of L. monocytogenes-positive samples (Table 2). Over the nine collection periods, winter had a consistently higher prevalence of L. monocytogenes than all other seasons, with the only exception being the summer of 2010. Farm 1 showed a significantly higher prevalence of L. monocytogenes than farm 2 (Table 2). The prevalence of L. monocytogenes was highest among water samples (48/174). All L. monocytogenes-positive water samples were from surface water (e.g., creek or pond water); none of the 28 samples from engineered water sources (e.g., municipal or well water) were positive for L. monocytogenes (Table 2).

Table 2.

Effect of farm, season, and sample type on frequency of samples positive for L. monocytogenes, Salmonella, and STEC found in produce preharvest environments

Factor (no. of samples)	Frequencya
	L. monocytogenes	Salmonella	STEC
Farm
1 (166)	39 (23)A	16 (10)A	1 (1)
2 (103)	5 (5)B	1 (1)B	3 (3)
3 (113)	13 (12)AB	3 (3)AB	7 (6)
4 (100)	14 (14)AB	5 (5)AB	1 (1)
5 (106)	17 (16)AB	2 (2)B	4 (4)
Season
Fall (136)	9 (7)B	3 (2)	4 (3)
Winter (125)	30 (24)A	6 (5)	2 (2)
Spring (134)	23 (17)AB	6 (4)	3 (2)
Summer (193)	26 (19)A	12 (9)	7 (5)
Sample type
Soil (178)	16 (9)B	4 (2)B	3 (2)
Drag swab (175)	15 (9)B	3 (2)B	5 (3)
Fecal (61)	9 (15)AB	4 (7)AB	4 (7)
Water (174)	48 (28)A	16 (9)A	4 (2)
Engineered (28)	0 (0)	0 (0)	0 (0)
Surface (146)	48 (33)	16 (11)	4 (3)

^aFrequency data represent the number of samples (percent). Superscript letters represent different statistical populations of values that are significantly different with P values of <0.016. Data with no letters represent values that are not significantly different.

Salmonella prevalence.

The prevalence of Salmonella across all samples was 4.6% (27/588). Farm and sample type were significantly associated with the frequency of Salmonella-positive samples (Table 2). While there was no significant seasonal association, Salmonella prevalence was greatest in the 2010 and 2011 summers (7.8 and 8.3%, respectively). Farm 1 showed a significantly higher prevalence of Salmonella than farms 2 and 5 (Table 2), possibly due to the comanagement of the produce operations on farm 1 with livestock operations located nearby. The prevalence of Salmonella was significantly higher in water samples (16/174) than soil and drag swab samples (4/178 and 3/175, respectively) but similar to that in fecal samples (4/61) (Table 2). All of the 16 Salmonella-positive water samples originated from surface water (Table 2).

STEC prevalence.

The prevalence of STEC was 2.7% (16/588) across all samples. Four samples tested positive for E. coli O157:H7, including a (i) drag swab sample from a pepper field, (ii) a drag swab sample from a sweet corn field, (iii) a water sample from a drainage ditch, and (iv) a water sample from a creek. None of the factors (e.g., farm, season, and sample type) were shown to have a significant association with the frequency of STEC-positive samples (Table 2). Similar to findings for L. monocytogenes and Salmonella, all four STEC-positive water samples were from surface water (Table 2).

L. monocytogenes diversity.

A total of 107 L. monocytogenes isolates were obtained from the collection of 88 environmental samples in which this pathogen was detected. Alignment of sigB nucleotide sequences for the 107 isolates showed 12 different sigB allelic types. Allelic types belonged to L. monocytogenes lineages I, II, and IIIa (6, 5, and 1 allelic types, respectively). There was a high diversity of L. monocytogenes allelic types among farms, seasons, and sample types (D = 0.80, 0.78, and 0.85, respectively).

There were four cases of repeat isolation, which was defined as the same sigB allelic type being isolated three or more times from the same sample site over time. These isolates were subtyped further using PFGE with restriction enzymes AscI and ApaI (Fig. 1). Analysis of L. monocytogenes PFGE showed multiple PFGE patterns for three of the four cases of repeat isolation; however, one case showed identical PFGE patterns for 3 of the 4 isolates obtained from the same water sample site.

Fig 1.

AscI and ApaI PFGE patterns of the four repeat isolation cases of L. monocytogenes. The four cases of repeat isolation are as follows: FSL-S10-009, -084, -366 and -598; FSL-S10-020, -086, and -601; FSL-S10-298, -604, and -1591; and FSL-S10-306, -335, -609, -1363, and -1490. In one case (top three PFGE patterns), three of four isolates (FSL-S10-009, -084, and- 366) have identical PFGE patterns. Band sizes (kb) are displayed at the top of the PFGE pattern images. The PFGE pattern order displayed is the result of sample site (manually ordered within BioNumerics).

Salmonella diversity.

Serotyping and PFGE were conducted on 1 representative isolate from each isolation scheme in 26 out of 27 positive samples for a total of 57 Salmonella isolates. No isolate from one Salmonella sample was available for typing, because preservation failed. One of the 26 available samples yielded a different PFGE type under the four isolation schemes. The two PFGE types were confirmed to be Salmonella serovars Newport and Thompson. All other isolates from the isolation schemes had identical PFGE types within a sample. The remaining 25 Salmonella-positive samples contained Salmonella serotypes Cerro (10 samples), Newport (5 samples), Thompson (4 samples), Give (2 samples), IV 40:z4,z32:− (2 samples), Typhimurium (1 sample), and I 6,8:i:− (1 sample) (Fig. 2). The 7 Salmonella serotypes corresponded to 11 different PFGE types (Fig. 2). Overall, there was a high level of diversity among Salmonella serotypes and PFGE types in the produce preharvest environment (for serotype, D = 0.84; for PFGE type, D = 0.80). Salmonella Newport was isolated from two fecal samples and one soil sample from the same field on farm 4 (Fig. 2).

Fig 2.

XbaI PFGE patterns of the 27 Salmonella isolates representing the 26 Salmonella-positive samples available for typing. One isolate per isolation scheme was typed; only one representative PFGE pattern is shown. One sample yielded two Salmonella PFGE patterns from the four isolation schemes, which represented Salmonella Newport and Thompson (FSL-S10-1570 and -1574). One case of repeat isolation, Salmonella Cerro from a surface water sample site, was identified (FSL S10-550, -1253, and -1411; depicted with boxes and bold text). Band sizes (kb) are displayed at the top of the PFGE pattern images. The PFGE pattern order displayed is the result of BioNumerics similarity analyses using UPGMA and the Dice correlation coefficient with a maximum space tolerance of 1.5%.

Repeat isolation of a Salmonella serotype was also observed. Salmonella Cerro was isolated from a water source three times during the nine collection periods (Fig. 2). The water sample was collected from a creek that was across the road from a field on farm 1. Salmonella Cerro is highly clonal, and this particular PFGE pattern matches 89% of the Salmonella Cerro PFGE patterns (55, 56).

STEC diversity.

Serotyping was conducted to further characterize the 16 STEC isolates. Six distinct O and H serotype results were observed. Serotypes O157:H7 (4/16 isolates) and O8:H19 (4/16 isolates) represented half of the 16 STEC isolates. Additional serotypes identified were O26:H11 (1/16 isolates), O−:H− (2/16 isolates), OX25:H11 (3/16 isolates), and O91:H49 (2/16 isolates).

Classification of high- and low-prevalence samples.

CT models were fit using sample presence/absence data in order to further explore the environmental and topographic variables that were associated with the detection of L. monocytogenes and Salmonella at smaller scales of variation than farm, season, or sample type (Fig. 3 and 4). CTs start with a root node containing all samples and recursively split sample sites by minimizing the mixture between positive and negative environmental samples for the selected food-borne pathogen. CTs often determine multiple possible rules useful for splitting samples (Fig. 3 and 4). Primary splits exhibited the best improvement score for dividing positive and negative samples into separate nodes; the rule with the second best improvement score was considered a competitor against the primary rule, except in cases where this rule was informationally redundant. In these cases, the next best competitor was selected for display (see Text S2 and S3 in the supplemental material). Surrogate rules represent the predictor that best correlates to the primary rule for the split, and they are used by the algorithm to fill in missing data for the primary rule. These surrogates mimic the primary rule and produce a split with a similar division of positive and negative samples in daughter nodes. No CT was developed for STEC because the trees produced only a root node.

Fig 3.

CT dividing L. monocytogenes environmental samples on the basis of remotely sensed topographical, edaphic, and meteorological data. On the top of each node there is a rule used for partitioning samples into homogeneous subsets. Primary rules are those used to make the depicted split. Competitor rules represent the rule with the second best improvement score. Surrogate rules mimic the primary rule and produce a split with a similar division of cultured positive and negative samples in daughter nodes. Percentages of cultured positive samples are displayed in each node. Rules partition toward the left-hand daughter node. Left-hand daughter nodes are enriched for negative samples, and right-hand daughter nodes are enriched for positive samples. Abbreviations: L, number of samples partitioned in the left daughter node; R, number of samples partitioned in the right daughter node, N, number of cultured negative samples; and P, number of cultured positive samples. Text S2 in the supplemental material provides a full summary of the CT. Briefly, Eigenvector_1_Temporal was an axis negatively correlated with temperature data, and splits based on this rule indicated that cooler temperatures are more likely to predict a pathogen-positive sample. Eigenvector_1_Spatial was an axis that was positively correlated with available water storage and organic matter values for soils. Locations with higher values on this eigenvector were likely to be pathogen positive.

Fig 4.

CT dividing Salmonella environmental samples on the basis of remotely sensed topographical and meteorological data. On the top of each node there is a rule used for partitioning samples into homogeneous subsets. Primary rules are those used to make the depicted split. Competitor rules represent the rule with the second best improvement score. Surrogate rules mimic the primary rule and produce a split with a similar division of cultured positive and negative samples in daughter nodes. Percentages of cultured positive samples are displayed in each node. Rules partition to the left-hand daughter node. Left-hand daughter nodes are enriched for negative samples, and right-hand daughter nodes are enriched for positive samples. Abbreviations: L, number of samples partitioned in the left daughter node; R, number of samples partitioned in the right daughter node; N, number of cultured negative samples; and P, number of cultured positive samples. Text S3 in the supplemental material provides a full summary of the CT. Briefly, Eigenvector_1_Spatial was positively correlated with soil moisture, and the primary split based on this variable indicated that soils below the maximum soil moisture in our database were likely to yield a positive sample. Eigenvector_2_Temporal was positively correlated with precipitation data and indicated that when measurable precipitation occurred within 3 days prior to sampling, a positive sample was more likely. Farm was determined to be an important variable, though this split appeared lower in the tree than soil characteristics and precipitation, indicating that the last two variables have larger-scale effects on the detection of Salmonella.

The L. monocytogenes CT that gave negative samples one-quarter the weight of positive samples resulted in the lowest relative cross-validation error at 0.65 (Fig. 3). The CT determined that the prevalence of L. monocytogenes in samples collected within 37.5 m of mapped waterways was 39% (29/74). All 74 L. monocytogenes samples within 37.5 m of mapped waterways were surface water samples. All L. monocytogenes-positive terrestrial samples (n = 40) were located farther than 37.5 m from mapped waterways. In sample locations ≥37.5 m from water, the eigenvector describing temperature variables split samples such that temperatures that were lower than approximately 2°C below average had 21% prevalence, but samples from warmer temperatures had only 7% prevalence. Using the remotely sensed average temperature over 5 days prior to sampling, the split rule (Eigenvector_1_Temporal < 1.483) corresponded to <14°C in summer, <10°C in spring, or <5°C in winter. The eigenvector for soil properties (Eigenvector_1_Spatial in Fig. 3) included available water storage and soil organic matter. Use of predicted available water storage from the SSURGO database as a representative value revealed that soils with available water storage of >4 cm at depths from 0 to 25 cm yielded samples with a 31% prevalence, whereas the prevalence was 10% in less moist soils. Proximity to pastures was also identified as an important factor in the prediction of L. monocytogenes prevalence. Moist terrestrial soil locations sampled at cooler temperatures within 62.5 m of a pasture had 50% L. monocytogenes prevalence (n = 25/50), whereas the L. monocytogenes prevalence in similar sample locations further than 62.5 m from pasture-class land areas was 7.5% (3/40). Land use classes were highly interspersed in the areas surrounding sampled farms, so the sample locations meeting the criteria for 50% L. monocytogenes prevalence occurred on four of five sampled farms, indicating that the proximity-to-pastures rule was not the product of bias due to the fact that farm 1 had a higher prevalence and shorter distances to pasture than the four others. A terminal node for proximity to urban development or roads (i.e., impervious surface coverage) was also identified. Locations within 9.5 m of an impervious surface had a predicted L. monocytogenes prevalence of 20%, whereas locations farther than 9.5 m had a prevalence of 5% (Fig. 3).

The Salmonella CT that gave negative samples 1/20 the weight of positive samples yielded the lowest relative cross-validation error, 0.67, across all our attempted weighting schemes (Fig. 4). Drainage class was identified as the most important factor delineating locations of high or low Salmonella prevalence. A location where drainage is classified as very poor, somewhat poor, poor, and somewhat drained was determined to have a higher Salmonella prevalence (9%) than a location where drainage is classified as moderately well drained and well drained (1.2%). After the tree determined that poorly drained soils contained more Salmonella, the algorithm then produced a rule indicating that soils near the upper limit of available water storage were more likely to be negative. Salmonella was less likely to occur in soils with available water storage (at 0 to 25 cm) of 10 cm, which was the maximum value in the soil database. The second temporal eigenvector, which described temporal variation in precipitation, formed another split. In areas with poorly drained soils, Salmonella was more prevalent (12%) when measurable precipitation occurred within 3 days prior to sampling.

Geospatial prediction of terrestrial L. monocytogenes reservoirs.

The CT contained topographic and soil property rules that might be useful to map the locations of environmental reservoirs of L. monocytogenes. In order to explore the usefulness of the CT rules, a map was generated to represent a CT-based hypothesis about the locations and spatial extents of L. monocytogenes environmental reservoirs in a 9,024-ha mixed-land-cover area in central New York State (Fig. 5). We hypothesize that these locations can harbor L. monocytogenes within croplands and are more likely to be positive when sampled than other parts of produce fields. Three classes of environmental reservoirs were extracted to produce this map: (i) areas within 38 m of mapped surface water, (ii) areas outside class (i) with soil available water storage (AWS) of >4.0 cm in the top 25 cm and within 62 m of mapped pasture land cover, and (iii) areas with AWS of >4.0 cm in the top 25 cm but outside classes (i) and (ii). All classes were clustered spatially to the west of the main stream in the map, indicating that not all portions of the landscape provide equally good reservoir habitats for L. monocytogenes, because the properties of the soil east of the stream differed. The algorithm identified 221 cropland reservoirs within 38 m of water (Fig. 5, light blue; 45% positive samples) and had areas ranging from 0.01 to 4.79 ha (median, 0.07 ha). One hundred ninety-two reservoirs in soil areas with higher available water storage (Fig. 5, cyan; 7.5% positive samples) ranged from 0.1 to 16.2 ha in area (median, 0.6 ha). Two hundred thirty-five reservoirs within 62 m of pasture-class land cover and in soils with higher available water storage (Fig. 5, dark blue; 50% positive samples) ranged from 0.1 to 5.2 ha in area (median, 0.5 ha).

Fig 5.

Map predicting environmental locations and spatial extents of areas with increased L. monocytogenes prevalence based on CT results (Fig. 3). Percentages in the legend indicate the prevalence of L. monocytogenes from the CT analysis. Cropland polygons were isolated and were assigned values as reservoirs on the basis of three levels of the CT: (i) areas within 38 m of mapped surface water (light blue, 45%), (ii) areas outside class (i) with soil available water storage (AWS) of >4.0 cm in the top 25 cm and within 62 m of mapped pasture land cover (dark blue, 50%), and (iii) areas with AWS of >4.0 cm in the top 25 cm but outside classes (i) and (ii) with 7.5% prevalence (cyan). The color surface represents proximity to the nearest pasture-class land cover. Brown cropland polygons are not expected to contain significant L. monocytogenes reservoirs. Black areas are water features contained in the U.S. Geological Survey National Hydrography Dataset available at http://nhd.usgs.gov/data.html.

DISCUSSION

One of the largest practical challenges to ensuring produce safety is to optimize the expenditure of financial and labor resources such that the pathogen will be detected where it is a public health risk. While many aspects of pathogen biology in the environment have been the topic of research (13, 57), field data on the nature of environmental pathogen reservoirs in produce fields are lacking. As a result, farmers and food safety professionals have few data on which to base sampling schemes that are intended to detect food-borne pathogens in the preharvest environment. This study describes the environmental distribution of three food-borne pathogens at the farm and field scales. We propose that using these analyses provides a means to improve surveillance for food-borne pathogens in produce fields by describing environmental variables that constrain the prevalence of pathogens. These data may also be used to identify areas of high and low predicted pathogen prevalence within farms, enabling more informed decisions about the management of crops associated with food-borne disease outbreaks. While this study does not directly measure the risk of produce contamination, these data can be used to support the development of risk models.

The present research was conducted by collecting diverse sample types on privately operated farms. While this has the advantage of enabling the collection of data in farms that are faced with the management challenges and practical considerations of businesses, it had the disadvantage that sample collection was dictated in large part by convenience to the landowner, so the sampling design quickly became unbalanced and smaller than was originally planned. However, it is important to note that field ecology studies frequently feature unbalanced and undersampled designs (18). As an alternative to violating the assumptions of logistic regression analysis, classification trees are an ideal method to analyze such unbalanced and undersampled designs (19). The method makes a single, simple assumption about the distribution of presence/absence observations: that these data can be subdivided according to environmental data to maximize the homogeneity of the observed results.

The prevalence of key food-borne pathogens is higher in water and fecal samples.

Water has been identified as both a reservoir and a transmission pathway for food-borne pathogen contamination of produce (17, 58). Moreover, all three pathogens examined in this study are known to be common contaminants of agricultural watersheds (59–62). The range of L. monocytogenes and Salmonella prevalence in watersheds has been estimated to be from 6.4 to 62% and 6 to 80%, respectively, based on the region of study (62–66). A 7% prevalence of Salmonella was observed in water samples obtained from a produce-growing region in California (67). Similarly, in the participating NYS produce farms, we obtained the highest prevalence of L. monocytogenes and Salmonella from water samples, demonstrating that water sources are a potential pathogen reservoir.

However, we recognize that there may be a potential bias toward a higher prevalence of pathogens in water samples than soil samples in this study. The abundance of food-borne pathogens in the environment is expected to be low, so there may be a greater chance of detecting a pathogen in water samples because the pathogen may be distributed more uniformly in water than soil samples. The five subsamples of soil collected were pooled in one composite sample per field, which may contribute to a lower prevalence estimate. L. monocytogenes- and Salmonella-positive samples collected from on-site surface water sources were mostly from small waterways, no more than 3 m wide, as many of the sampled produce farms are surrounded by residential land. Previous research has shown that lower-order waterways often receive more direct agricultural drainage and runoff from livestock production environments than higher-order systems (60, 61).

Foraging wildlife may also contribute to contamination of fields, as demonstrated by isolation of Salmonella Newport from soil and wildlife (deer) feces from a single field. Wild and domestic animals are widely known to harbor food-borne pathogens (6, 67, 68). A review of Salmonella in wild and domesticated animals (68) determined the prevalence of Salmonella in reptiles (6 to 100%), poultry (0 to 60%), cattle (2 to 42%), swine (3.5 to 28%), rodents (1 to 15%), birds (3 to 13%), and domestic cats and dogs (1 to 5%), which reflects the potential for animals to be sources or vectors of preharvest contamination. An optimal solution for the comanagement of wildlife habitat, environmental quality, and food safety is the target of current produce safety research, and our research facilitates these efforts by describing the prevalence of pathogens in preharvest environments relative to potential wildlife habitats on the same landscape (69).

Soil properties and topographic features were identified as constraints on pathogen prevalence in produce fields.

The L. monocytogenes and Salmonella CTs both demonstrate that not all croplands have an equal risk of food-borne pathogen contamination. Soil characteristics and topographic variables corresponding to the proximity of sampled areas to other landscape types, including impervious surface (e.g., buildings, roadways), water, or pasture, were identified as factors (i.e., primary rules in the CT) for predicting locations containing pathogens. The primary split for both pathogen CTs was associated with water features, specifically, proximity to water and soil drainage class for L. monocytogenes and Salmonella, respectively. This finding further demonstrates the ability of on-site water sources to be potential reservoirs and transmission pathways for food-borne pathogen contamination in the preharvest environment (4, 17, 70).

Proximity to pastures was a factor influencing the likelihood of detecting an L. monocytogenes-positive sample. It is important to note that proximity to pastures was obtained through remotely sensed data, and pasture-class land cover can indicate active pasturages, livestock pens, and hay grass fields. It has been shown that livestock shedding and subsequent runoff of food-borne pathogens may be one of the major sources of preharvest contamination (17, 71, 72). Ruminants can shed significant numbers of L. monocytogenes while being asymptomatic and may release the pathogen into the environment (26, 56). A strong association between the prevalence of L. monocytogenes and proximity to cattle and dairy farms was shown in watersheds impacted in agricultural landscapes (27, 60).

Proximity to impervious surfaces was also identified as a factor for the classification of high- or low-L. monocytogenes-prevalence locations. Some wildlife carriers of food-borne disease, e.g., rodents and ground-nesting birds, use roadside ditches as nesting habitats and may enter croplands from ditches while foraging for food (73). This behavior, particularly in short-dispersal-distance species, may cause pathogen prevalence to be amplified in the edges between residential land and cropland. Impervious surfaces are constructed to remove excess water, usually into bordering drainage ditches, and these may channel fecal contamination. This contamination could subsequently be spread out of the ditch by local wildlife, heavy precipitation, or human activities.

Soil characteristics, specifically, available water storage, soil drainage class, or soil organic matter, were important factors in CTs generated from both L. monocytogenes and Salmonella prevalence data. Pathogen survival has been shown to increase in moist soils (74, 75). E. coli and Salmonella held in dry soil for 14 days demonstrated the ability for growth after the soil was moistened with sterile distilled deionized water (76). In addition, it was shown that E. coli and Salmonella persisted longer in moist soils than drier soils. Similarly, E. coli numbers were shown to decrease faster in dry soils than moist soils (77).

Weather plays a role in pathogen prevalence in the preharvest environment.

Meteorological factors were identified in both the L. monocytogenes and Salmonella CTs, indicating that recent temperature and precipitation can influence food-borne pathogens in the environment, which is consistent with previous findings (19, 62, 67, 78). L. monocytogenes was more frequently detected in cooler temperatures that were above freezing prior to sample collection in our study. Two previous studies (66, 79) observed a higher prevalence of L. monocytogenes in spring and winter-spring. Ivanek et al. (19) determined that fewer freeze-thaw cycles increased the detection of Listeria spp. in vegetation samples (19). L. monocytogenes can grow and survive in a wide range of temperatures, but freezing can have an inhibitory effect (80, 81). However, temperature was not found to be a factor in the CT for Salmonella. In contrast, other studies (62, 78) have identified temperature or season to be associated with the frequencies of Salmonella in environmental settings. It may be concluded that the ability to predict an increased likelihood of detecting a positive Salmonella sample, as for other species, may be dependent on the local ecology and agriculture practices of the specific location of the study.

Precipitation was identified as a factor influencing the detection of Salmonella-positive samples in our study. It has been suggested that heavy rain and storm events may aid in the transport of pathogens (82, 83) and potentially lead to higher loads of bacteria in the water, along with sediment (84, 85). Food-borne pathogens can survive in sediments for substantial periods (86). High water flow rates have also been observed to influence pathogen incidence levels in the water and may transport pathogens up to 32 km (87). Such long transport distances in waterways may have important implications for the diversity and source tracking of food-borne pathogens impacting preharvest environments (64, 71). It may be necessary to analyze the flux of pathogens from potential sources, like livestock pasture, through hydrologic connectivity networks in a dynamic framework that accounts for precipitation in order to accurately estimate the risk of Salmonella contamination in croplands where it is expected to result from contaminated water sources. The Soil and Water Assessment Tool (SWAT), while unsophisticated in its treatment of bacteria as simple particles, is one tool available to model this process.

Benefits to predicting the presence of food-borne pathogens in produce environments.

Transmission of food-borne pathogens to produce in the preharvest environment is a complex process, involving multiple vehicles that transport pathogens from sources (e.g., pasture areas) to sinks (e.g., moist soils in fields). The development of practical tools to predict the presence of pathogens in produce fields or the risk of produce contamination is further complicated by the fact that every produce farm contains a unique combination of spatial and temporal variables. These variables undoubtedly influence the ecology of food-borne pathogens in the environment and may influence the potential for product contamination. Indeed, the presented analysis indicates that, while the average prevalence of food-borne pathogens in randomly collected environmental samples is low, local prevalence can be significantly higher under specific combinations of environmental and local land use conditions. A component essential to developing a mechanistic approach to understanding food-borne pathogen transmission to produce is to characterize environmental reservoirs as favorable or unfavorable pathogen persistence sites. Since CT analysis generates concise rules to delineate pathogen-positive and -negative sites, application of the resulting rules to remotely sensed data about farm landscapes can enable the development of specific predictions about expected pathogen presence for any individual produce field. While the CTs presented here require further validation to determine their ultimate usefulness in produce farms, these models advance GIS-enabled modeling to predict the risk of produce contamination. Additionally, fully developed models of how pathogens disperse and persist in the preharvest environment may also permit the development of land management strategies to minimize produce contamination by possibly allowing growers to select crops for these sites that are less susceptible to contamination.