https://orcid.org/0000-0002-8730-1054
Figures
Abstract
Human exposure to mycotoxins is common and often severe in underregulated maize-based food systems. This study explored how monitoring of these systems could help to identify when and where outbreaks occur and inform potential mitigation efforts. Within a maize smallholder system in Kongwa District, Tanzania, we performed two food surveys of mycotoxin contamination at local grain mills, documenting high levels of aflatoxins and fumonisins in maize destined for human consumption. A farmer questionnaire documented diverse pre-harvest and post-harvest practices among smallholder farmers. We modeled maize aflatoxins and fumonisins as a function of diverse indicators of mycotoxin risk based on survey data, high-resolution geospatial environmental data (normalized difference vegetation index and soil quality), and proximal near-infrared spectroscopy. Interestingly, mixed linear models revealed that all data types explained some portion of variance in aflatoxin and fumonisin concentrations. Including all covariates, 2015 models explained 27.6% and 20.6% of variation in aflatoxin and fumonisin, and 2019 models explained 39.4% and 40.0% of variation in aflatoxin and fumonisin. This study demonstrates the value of using low-cost risk factors to model mycotoxins and provides a framework for designing and implementing mycotoxin monitoring within smallholder settings.
Citation: Stafstrom W, Ngure F, Mshanga J, Wells H, Nelson RJ, Mischler J (2025) Modeling maize aflatoxins and fumonisins in a Tanzanian smallholder system: Accounting for diverse risk factors improves mycotoxin models. PLoS ONE 20(1): e0316457. https://doi.org/10.1371/journal.pone.0316457
Editor: Ana Cyra dos Santos Lucas, Universidade Federal do Amazonas, BRAZIL
Received: July 26, 2024; Accepted: December 11, 2024; Published: January 13, 2025
Copyright: © 2025 Stafstrom et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: All data and code used in this paper are available in a public Github repository that has a Zenodo persistent identifier (DOI 10.5281/zenodo.14226520). Additional relevant data (the 2019 survey questionnaire, the 2015 and 2019 survey data, model results) are within the paper and its Supporting Information files.
Funding: Funding for this research was provided by the Mycotoxin Mitigation Trial (Grant No. OPP1155626) from the Bill & Melinda Gates Foundation (https://www.gatesfoundation.org/) (RJN). The McKnight Foundation Collaborative Crop Research Program (https://www.mcknight.org/) also supported this research (RJN). Key support for the 2019 survey was provided by the Cornell Atkinson Center for Sustainability Small Grants Program (https://www.atkinson.cornell.edu/), and by a travel grant from the Mario Einaudi Center for International Studies (https://einaudi.cornell.edu/) (WS). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing interests: The authors have declared that no competing interests exist.
Introduction
Ensuring safe and secure food systems is a global priority, and mycotoxin contamination is a significant obstacle to this goal. Mycotoxins are secondary metabolites produced by fungal plant pathogens that are toxic to humans and livestock. The negative effects caused by trace amounts of mycotoxins in staple foods are a significant challenge to food security, causing economic losses and burdens on public health [1, 2]. Globally, mycotoxins exceed regulatory limits in an estimated 25% of foods [3]. High-income countries employ resource-intensive regulatory structures and repurpose or eliminate foods exceeding established mycotoxin standards. These systems are largely effective but incur significant economic costs; in the United States of America, crop losses and mitigation efforts cost the economy an estimated $1.5 billion annually [4].
In resource-limited contexts, regulatory enforcement of mycotoxins is generally weak. The lack of effective regulation in many mycotoxin-prone environments means that chronic and episodic exposure to high levels of mycotoxins causes severe public health consequences. Liver cancer is associated with chronic aflatoxin consumption, in utero aflatoxin exposure increases rates of growth faltering, childhood stunting is associated with the presence fumonisin and aflatoxin in baby foods, and outbreaks of acute aflatoxicosis in Kenya and Tanzania have resulted in hundreds of deaths [5–12]. Frequently, one of the primary sources of mycotoxin-contaminated food is maize.
Maize is a key staple crop across many parts of sub-Saharan Africa (SSA) such as Tanzania, where it accounts for approximately forty percent of daily calories [13]. In Tanzania, milled maize is typically consumed as ugali, a stiff porridge, or it is combined with other milled grains and groundnuts to create a composite flour, lishe, that is used to make uji, a thinner porridge prepared primarily for children [14]. Consumption of ugali and uji exposes children and adults to numerous mycotoxins [15, 16]. Numerous mycotoxin-producing fungi infect and contaminate maize grain including, aflatoxin-producing Aspergillus flavus and A. parasiticus, fumonisin-producing Fusarium verticillioides and F. proliferatum, and deoxynivalenol-producing F. graminearum [17, 18]. Naturally occurring aflatoxins are classified as a Group 1 human carcinogen, and the most prevalent and toxic form of fumonisin, Fumonisin B1, is a Group 2B possible human carcinogen [19]. Maize is more susceptible to aflatoxin and fumonisin contamination than small grains like sorghums and millets [20, 21]. Mycotoxigenic fungi are adapted to different environments, but co-occurrence is frequently observed, particularly in warm and dry areas of SSA where both A. flavus and F. verticillioides thrive and produce aflatoxins and fumonisins, respectively [22, 23].
Maize is processed in SSA at a range of scales, from large-scale industrialized operations to small rural local mills [24]. Across most of SSA, mycotoxin regulations are largely unenforced, especially in rural subsistence contexts [25]. Smallholder farms are important maize producers in SSA, and these farms can be subsistence-only operations or, more commonly, they also interact with markets as buyers and/or sellers [26]. Smallholder farm systems are typically low-input and rainfed, which makes them more vulnerable to abiotic stressors like water and nutrient deficiencies [27]. With climate change, extreme weather events like drought are predicted (and observed) to be increasingly common in SSA and are projected to reduce maize yields and increase the risk of mycotoxin contamination [28–31].
Environmental and cultural factors intersect to make maize consumers in smallholder systems especially vulnerable to mycotoxin exposure [32]. Extreme hot and dry weather during the growing season stresses developing maize plants, making them more susceptible to infection by mycotoxigenic fungi, and thus at increased risk of mycotoxin contamination [33, 34]. Hotter weather provides the optimal growth temperatures for these A. flavus (32°C) and F. verticillioides (30°C) [35–37]. Also, hot and dry weather during flowering and grain fill periods of maize development facilitates opportunistic A. flavus infection and is associated with increased aflatoxin contamination, while warm and wet conditions later in the season provide an ideal environment for established fungal pathogens to grow [38]. Warmer temperatures during flowering were also associated with increased fumonisin contamination in maize [39]. Developing the capacity to predict mycotoxin accumulation in these production systems requires integrating knowledge of geospatial mycotoxin dynamics with local information on diverse mycotoxin risk factors.
Environmental datasets with high spatial and temporal resolution are increasingly useful for agricultural applications due to improved satellite imagery (e.g., Sentinel-2), more powerful statistical modeling techniques (e.g., machine learning methods), and collaborative efforts to generate agriculturally-relevant metrics (e.g., Famine Early Warning Systems Network (FEWSNET)) [40–43]. Such datasets enable enhanced monitoring of pre-harvest mycotoxin risk factors in rural settings. Combined with georeferenced mycotoxin data, they are vital to developing statistical models that can predict the distribution of mycotoxins across a landscape. For example, in Georgia, USA, weather station data was combined with soil drainage, normalized difference vegetation index (NDVI), and thermal infrared data to model historical aflatoxin contamination in maize [44]. In SSA, NDVI imagery predicted groundnut aflatoxins in Mali, while in Kenya, a combination of NDVI, rainfall, and soil data enabled modeling of maize aflatoxins [45, 46].
In each of these cases, the timing of key indicators of plant stress was also a critical factor, as pre-harvest fungal infection and mycotoxin accumulation are most likely to occur during the flowering and grain fill growth periods [47, 48]. Defining these critical periods helps to improve prediction models of maize aflatoxins in inoculated field trials [36, 49, 50]. Although monitoring the dynamics of pre-harvest environmental factors can be effective and scalable, it cannot account for all drivers of mycotoxin contamination of food, both because mycotoxins can accumulate post-harvest and because many factors are determined at the farm or household scale.
Smallholder maize systems in SSA are complex and diverse in terms of crop genetics, soil conditions, management techniques, processing decisions, and post-harvest handling. Many of these factors influence mycotoxin contamination, and some of them are managed or influenced by actions of farmers and others. Locally sourced information on risk factors and behaviors specific to a farm or community can improve mycotoxin modeling that relies on environmental datasets, and can also help inform better mycotoxin management, including responses to outbreaks. Data on key variables may be ascertained with direct input from the farmers or communities themselves [51]. Indeed, farmer-reported data (variety, fertilizer use, planting, and harvest dates) was key to modeling maize fumonisins in northern Italy [52].
In SSA, information provided by farmers and consumers about production, storage, and processing could also prove useful in modeling mycotoxins. For instance, the application of nitrogen fertilizer reduced maize aflatoxins in Kenyan research trials, suggesting the need to obtain information from farmers about their input use [53]. The post-harvest practice of sorting out moldy, fragmented, or discolored kernels reduces maize fumonisins [22, 54, 55]. In maize processing, dehulling maize grain reduces fumonisin levels, as F. verticillioides growth is mostly restricted to the outer layers of the grain (the aleurone and pericarp), which are removed during this process [56, 57]. In Tanzania, maize is typically processed as either dona flour (whole-grain) or sembe flour (dehulled), and in certain regions it is also common to make kiwerege flour (soaking dehulled maize in water before milling) [14, 58]. Information about crop management, storage and processing are thus important for assessing mycotoxin risk.
Developing powerful empirical mycotoxin models depends on acquiring a large amount of high-quality mycotoxin data. Mycotoxin assays (e.g. lateral flow devices, enzyme-linked immunosorbent assays (ELISA), etc.) are necessary for any monitoring effort, but their cost and infrastructure requirements often preclude their implementation in low-resource settings [59]. Spectral assays show promise for being an affordable and portable option, particularly in the rural SSA context [60–62]. Although there are tradeoffs in accuracy and a subset of the samples must be tested with validated assays for model training, the ability to use a spectral-based assay to cheaply estimate mycotoxin levels in numerous samples in the field could bolster landscape-wide mycotoxin prediction.
While most prior mycotoxin modeling studies have relied on environmental covariates [46, 63], we hypothesized that a more holistic approach that integrates information on local indicators of mycotoxin risk could be more effective at identifying at-risk areas and guiding mitigation efforts (e.g. breeding resistant varieties, implementing grain sorting practices, improved post-harvest storage methods). This study used a household and grain mill mycotoxin survey in central Tanzania as a platform to test a feasible and multifaceted approach for modeling maize aflatoxins and fumonisins in a smallholder farming system. This approach was based on sampling at local nodes of maize processing, monitoring high-resolution environmental risk factors, surveying farmers’ pre- and post-harvest practices, and using a low-cost spectral device to analyze grain. This effort could provide a framework for mycotoxin monitoring within a low-resource setting, and, ultimately, inform and support mycotoxin mitigation efforts.
Materials and methods
Study site
All food surveys occurred in Kongwa District, Tanzania, which has an area of 4,041 square kilometers and a semi-arid agro-pastoral farming system. Among the district’s 309,973 residents, many are smallholder farmers who grow maize as a staple crop in a high temperature and low rainfall (average of 700 mm of rainfall per year) environment [64]. The main crops by cultivated area are maize, millet, sunflower, groundnuts, cassava, and sorghum [64]. This district was also the site of a randomized control trial for testing the effect of aflatoxin exposure on childhood stunting [65].
2015 food survey
Sampling.
The first food survey was conducted from November 10–20, 2015 in 19 villages across Kongwa District. Village leaders led the survey team to households and grain mills where samples of locally processed maize flour samples were obtained. In households where maize flour samples were not available a random sample of whole kernel maize was drawn from pantry stores. Latitude and longitude coordinates of grain mills were acquired in the WGS84 reference coordinate system by using the ODK Collect app, and households were assigned the coordinates of their local grain mill (Get ODK Inc.). Information was collected on the type of maize flour and the source of the maize (homegrown or purchased).
Mycotoxin assays.
A total of 128 maize samples were collected during this food survey. Total aflatoxins and fumonisins from milled samples were quantified in the field using their respective Neogen Reveal Q+ lateral flow strips. For whole kernel samples, milling and quantification were performed within one week of collection at Nelson Mandela African Institute of Science and Technology (NM-AIST) in Arusha, Tanzania. For aflatoxin and fumonisin extractions, a random 10 g sub-sample of milled maize was placed in 50 ml of 65% ethanol, vortexed for three minutes, and filtered with Whatman #1 filter paper. Then, extracted fumonisins and aflatoxins were assayed according to Neogen kit protocols. Aflatoxins or fumonisins were quantified from their lateral flow strips using the Lab-on-Mobile-Device platform (Mobile Assay Inc.) [66]. Samples that exceeded the Neogen Reveal Q+ limits of detection (2 μg/kg for aflatoxin, and 300 μg/kg for fumonisin) were serially diluted and retested.
2019 food survey and farmer questionnaire
Sampling and questionnaires.
For the 2019 food survey, at least three villages were visited within each ward in Kongwa District. At grain mills, participants were recruited, and interviews were conducted from May 17, 2019 to June 6, 2019. Prior to visiting grain mills, local village leadership was consulted, and they introduced the survey team to the grain mill proprietor. At each grain mill, ODK Collect was used to acquire latitude-longitude coordinates (WGS 84 datum), and then convenience sampling was conducted with the aim of interviewing 3–10 patrons per mill (Get ODK Inc.). Grain mill patrons were informed of the survey’s aims and, if eligible, a brief interview and to a sample of their processed maize meal. The following criteria were applied to identify survey participants: 1) the patron was 18 years of age or older, 2) the maize originated from their local farm, 3) the patron had knowledge on how the maize was grown and how it was handled post-harvest, and 4) the maize was harvested from the current growing season.
Respondents were interviewed about the growing conditions of their maize, their agronomic decisions, and their storage and processing choices, and their responses were recorded in the ODK collect app (S1 Appendix, Get ODK Inc.). Following the interview, the customer’s milled maize was sampled (approximately 300 g), and customers were compensated for their time and maize sample with 2,000 Tanzanian shillings (~US$1). In total, surveyors visited 90 grain mills in 38 villages and conducted 342 surveys, collecting the corresponding number of maize samples.
Mycotoxin assays.
Milled maize samples were transported to the NM-AIST in Arusha, Tanzania. Samples that were improperly labelled (n = 36) were discarded. Mycotoxin extraction and quantification were conducted on 306 samples according to protocols for total aflatoxins and total fumonisins ELISA kits (Helica Biosystems Inc.).
For each comminuted sample, random subsamples of 9 g and 17 g were measured in separate 50 ml Falcon tubes for aflatoxin and fumonisin extractions, respectively. For aflatoxin extraction, 45 ml of 70% methanol was added, and for fumonisin extraction, 34 ml of 90% methanol was added. Samples were agitated on a shaker for 30 minutes, centrifuged at 10,000 rpm for two minutes, filtered through Whatman #1 filter paper, and 2 ml subsamples of the extracts were stored at 4°C in 2 ml Eppendorf tubes. Following sample extraction, ELISAs for total aflatoxin and total fumonisins were performed according to Helica protocols for each toxin. A BioTek Synergy 2 plate reader with Gen5 software was used to measure the optical density (OD, absorbance at 450 nm) of each well (BioTek Instruments Inc.). A standard curve was fitted to the standards’ OD values, and total aflatoxin or total fumonisin concentrations for each sample were interpolated. If a sample exceeded the ELISA limit of detection (20 μg/kg for aflatoxin and 6 mg/kg for fumonisin), it was diluted and rerun on a new plate. A common sample was assayed on each plate and the between-plate coefficient of variation (CV) threshold was 15%, and duplicate samples within plates were held to a 10% CV threshold. Total aflatoxin and fumonisin values were positively skewed, and a log10-transformation was applied to each prior to inclusion in statistical models (S1 and S2 Tables).
NIR analysis.
Before mycotoxin extraction, a random subsample of each milled maize sample was scanned three times with a SCiO mini NIR (near-infrared) spectrometer (Consumer Physics). This spectrometer measured reflectance from 740–1070 nm at 1 nm intervals, but because of noise at either end of this spectrum, the range was trimmed to 766–1044 nm. For each sample, the mean reflectance of the three replicate measurements was calculated (S2 Table). To account for light scattering and variability in sensor-sample distance, pre-processing techniques were applied to the mean reflectance values. First, the wavelengths were smoothed using a Savitzky-Golay transformation (differentiation order = 1, polynomial order = 1, window size = 13), and then these transformed wavelengths were normalized using a standard normal variate transformation. A principal components analysis (PCA) was performed on these 279 pre-processed wavelengths.
The NIR data were also used to test whether it could identify different flour types. A random forest classification model (number of trees = 500, variables sampled at each split = 16) was built using all 279 transformed wavelengths as predictors and the three maize types (dona, sembe, and kiwerege) as response variables. The importance of each wavelength was calculated as the percent increase in mean square error (MSE) for permuted out-of-bag data.
Geospatial datasets acquisition and processing
The elevation of each household or grain mill’s coordinates was acquired from AWS terrain tiles [67]. Environmental datasets were identified based on several criteria including: high spatial resolution (<1 km2), previously identified as a pre-harvest risk factor for aflatoxin or fumonisin contamination, and, when applicable, high temporal resolution (≤1 dekad or 10 days). Two datasets that satisfied these criteria had 250 m2 spatial resolution: 1) dekadal temporally-smoothed NDVI from FEWSNET, and 2) soil attributes at depth intervals in SoilGrids250m [41, 42]. Based on reported planting and harvest dates in 2019, the maize growing season was defined as spanning December dekad 1 to May dekad 3. For the 2014–2015 and 2018–2019 growing seasons, NDVI values were acquired from the USGS FEWSNET data portal for these 18 dekads.
To ensure that non-farmland signals (e.g. those from inhabited areas, rangeland, bodies of water, brush or forest, etc.) were not included in the analysis, cropland was classified using representative wet season imagery. For each year-dataset combination, non-cropland pixels were masked and excluded from further analyses.
We defined a millshed as the catchment area that was likely to contain farms whose harvests were brought to a given mill. These millsheds were used to capture local environmental variation from high resolution datasets. For the 2015 dataset, NDVI and soil values were extracted from a 5 km radius buffer surrounding each georeferenced mill and averaged (S1 Table). Household samples were assigned the value of their nearest millshed.
In 2019, millsheds were calculated for individual samples by converting reported farm-to-mill walking times into estimated distances from the grain mill. To find estimated distances the reported time was divided by a 3 km/hour walking speed. Further restrictions were a 0.5 km minimum and 10 km maximum on walking distances. Any unreported distances were assigned the mean walking distance. These distances were used to create a distance buffer radius from the grain mill for extracting and averaging soil and NDVI values for each sample (S2 Table).
A PCA was performed on each year-dataset combination; in total, four separate PCAs were run on the extracted and buffer-averaged values from the 2015 and 2019 NDVI and soil datasets. The number of PCs in models was determined by including those that explained up to 90% of cumulative variation for a given dataset.
2015 mycotoxin models
To assess the influence of survey and environmental factors on maize mycotoxins from the initial 2015 survey, multiple linear regression models were constructed using these potential mycotoxin risk factors as covariates. The response for each model was log10-transformed aflatoxin or log10-transformed fumonisin values, and the covariates included flour type, source (homegrown vs. purchased), household elevation, the first five principal components of all eleven of soilGrids250m attributes, and the first three principal components of the entire growing season’s dekadal NDVI values. Then, to test whether environmental factors were more effective at modeling mycotoxins in homegrown maize compared to purchased maize, the same models (with the exception of “source”) were constructed using only homegrown or only purchased samples.
2019 mycotoxin models
One-way analysis of variance (ANOVA) was used to test the effect of maize flour type on transformed aflatoxin or fumonisin levels within each year. If significant, post-hoc pairwise t-tests between each group were performed and the Holms method was used to adjust for multiple comparisons.
Mixed linear models were used to model aflatoxins and fumonisins. First, a mixed linear model was built using the first five NIR PCs, the first three NDVI PCs, the first five soil PCs as fixed effects and mill village as a random effect. Before constructing linear mixed models, the mean at each of the 88 grain mills was calculated for all attributes of the 306 samples, and these values were used for the response and predictor variables.
Software and data analysis
Data processing and statistical analysis were conducted in R version 4.3.1 [68]. For NIR data, the prospectR package was used to pre-process NIR wavelengths [69]. For environmental data processing, the raster package was used [70]. For linear mixed models, the lme4 package was used to construct the models, the lmerTest package was used to identify significant covariates, and the MuMin package was used to calculate conditional and marginal coefficients of determination [71–73]. Random forest modeling and variable importance calculations were performed using the randomForest package in R [74]. Elevation data for each location were acquired using the elevatr package in R [67]. Data were visualized using the ggplot2 and ggpubr R packages [75, 76].
Ethical considerations
Cornell University’s Institutional Review Board provided the ethical approval for farmer questionnaires (Protocol ID# 1809008284). As documented in the approval’s standard operating procedure for this study, individual farmer interviews were conducted in Swahili. Research team members read a script to potential participants explaining the purpose of the study and the criteria for participant inclusion, and, if they qualified, their consent for a sample and a brief interview was politely solicited. Oral informed consent was documented by the lead member of the research team on the data form and witnessed by the second member of the team. No identifying information was recorded, and data were analyzed anonymously. Administrative permission to visit local grain mills was received from the Kongwa District Executive Director.
Results
Kongwa District survey insights
Many preliminary indicators suggested that the residents of Kongwa District would be at high risk of mycotoxin exposure. In 2019, we characterized the local context with a farmer questionnaire, both to better understand mycotoxin risk factors and to better envisage management options. In 2015, the food survey accounted only for whether the maize was homegrown (59.8%) or purchased (40.2%) and the type of maize flour (dona = 37.6%, sembe = 53.0%, kiwerege = 9.4%). The 2019 food survey only sampled homegrown maize and used a more extensive questionnaire to gain insights into pre-harvest factors, agronomic decisions, post-harvest handling, and processing decisions (Table 1). Kongwa residents produce maize for either subsistence use and/or to sell at the market. Self-provisioned maize was commonly processed at grain mills that were located within walking distance of their farms (mean = 77.0 minutes and median = 60.0 minutes).
Maize was planted as early as the beginning of October 2018, but most farmers planted in December 2018 or January 2019. Farmers explained that their planting decisions were guided by local rainfall patterns and their individual risk tolerance. Planting immediately after the initial rains in early December was described as a higher risk decision, while those with moderate risk tolerance planted in late December, and the most risk averse waited until January to plant after rains had definitively established. Season length (harvest dekad-planting dekad) had a median of 14 dekads (~140 days), but maize was in the field for as little as 8 dekads and as long 22 dekads. These season lengths can be explained in part by farmers allowing their maize to dry in the field and harvesting portions of their field as needed.
Farmers used very few inputs. Most seed had been saved from the previous year, fertilizer use was rare, and there were no reports of irrigation. In terms of pre-harvest conditions, nearly all farmers reported that their maize experienced drought-like conditions, especially in the middle of the season (February-March). Though there is commonly a short break in rains in January or February, the intensity and duration were described as more severe than normal in the 2018–2019 growing season. Most farmers allowed their maize to mature and dry in the field before harvesting, and they reported that this matured and dried maize was rained on. Farmers grew their maize in diverse cropping systems with two to six other crops intercropped with their maize. Leguminous and oil crops were the most common classes of crops, with the cash crop sunflower being the most common choice for intercropping.
Following harvesting and drying, maize was typically stored in polypropylene bags, though some maize was brought directly to the mill for processing. The vast majority of those interviewed winnowed and hand-sorted their maize prior to milling, but we observed that the intensity of hand-sorting varied (e.g. from quickly removing a few rocks and cobs to painstakingly removing all abnormal kernels).
Respondents processed their maize using different methods: 1) milling whole grain (dona), 2) dehulling and milling (sembe), and 3) dehulling, soaking in water for at least one night, and then milling (kiwerege). These processing choices were influenced by a variety of factors including food security, labor required, food quality, and personal preference. For instance, some would opt to go to certain mills based on their preference for how much bran and germ would be removed in the dehulling process.
Kongwa District environmental insights
To further contextualize mycotoxin risk factors in Kongwa District, we characterized the environmental conditions under which the maize was likely grown. Sampling occurred across the entirety of Kongwa District where the elevation of surveyed grain mills had a mean and median of 1,214 and 1,224 masl and a minimum and maximum of 904 and 1,492 masl (Fig 1).
Elevation (meters above sea level) map of Kongwa District with 32 millsheds sampled in 2015 (red dots) and 90 millsheds sampled in 2019 (black dots).
NDVI values extracted from millshed buffers illustrated the dynamics of vegetation growth across the maize growing seasons. For instance, maximum mean millshed NDVI values were observed during the final two dekads of March, which roughly indicates the flowering time of maize, as this demarcates a transition between maize’s vegetative and reproductive growth phases (Fig 2). Comparing the two seasons, NDVI values diverged halfway through the growing season; 2015 NDVI values were higher than those in 2019 from the end of February to mid-May. The 2019 NDVI values were consistent with farmers’ characterization of the middle of the growing season as having little to no rainfall; following this period (February to March 2018), there was a flattening of NDVI, which suggested that plant health was negatively affected by the reported lack of rain. Relative to the norm (median NDVI between 2003–2017), both seasons had depressed NDVI from the middle to end of the growing season (March-May), and this plant health deficit was particularly pronounced in the 2018–2019 growing season. This indicated that maize crops were stressed during the maize flowering and grain fill periods.
(A) Millshed mean NDVI dekadal values across the 2014–2015 (n = 32) and 2018–2019 (n = 90) maize growing seasons in Kongwa District (black lines = dekad mean). At each growing season dekad, a two-sample t-test was run to test for differences between NDVI values in 2015 and 2019, and p-values were adjusted for multiple comparisons with the Holm method (ns p>0.05, * p<0.05, ** p<0.01, *** p<0.001, **** <0.0001). (B) Millshed mean NDVI values as a percentage of the median NDVI from 2003–2017 (black lines = dekad mean, dashed line = equal to median of 2003–2017). At each dekad, a two-sample t-test was run to test for differences between NDVI percentage median values in 2015 and 2019, and p-values were adjusted with the Holm method (ns p>0.05, * p<0.05, ** p<0.01, *** p<0.001, **** <0.0001).
Soil plays a vital role in supporting plant health and can serve as a buffer against potential stressors. Whereas a pre-harvest factor like NDVI varies within seasons and across years, available soil data were static relative to time. Assessment of a soil factor that has been negatively associated with mycotoxin contamination, soil organic carbon (SOC), showed that millsheds’ values did not differ significantly across years (Fig 3).
Millshed mean soil organic carbon (SOC) in the fine earth fraction across depths for the 2014–2015 and 2018–2019 maize growing seasons (black lines = depth mean). At each soil depth, a two-sample t-test was performed to test for differences in SOC means in 2015 and 2019, and p-values were adjusted with the Holm method (ns p>0.05).
Kongwa District NIR insights
Using NIR measurements of maize flour samples, random forest classification was able to classify flour type with an out-of-bag error rate of 10.3%. For the different flour types, error rates were 12.0% for dona, 7.8% for sembe, and 31.3% for kiwerege, though low accuracy for kiwerege identification was likely influenced by its small number of samples (n = 16). Measuring percent increase MSE showed that 773 and 805 nm were the most important NIR wavelengths for flour type identification (Fig 4).
Percent increase in mean square error (MSE) for all near-infrared wavelengths in a random forest classification model for maize flour type.
Mycotoxins and surveyed risk factors
In 2015, milled maize samples collected from grain mills and household stores had a geometric mean (95% CI) of 11.6 μg/kg (9.1–14.7 μg/kg) for maize aflatoxins, and 0.094 mg/kg (0.071–0.12 mg/kg) for maize fumonisins. Maize mycotoxin contamination was more severe in 2019 than in 2015, with a geometric mean (95% CI) of 29.9 μg/kg (25.0–35.7 μg/kg) for total aflatoxins and 0.80 mg/kg (0.70–0.99 mg/kg) for total fumonisins (Fig 5). In relation to the East African Community’s standards for maize flour, a substantial percentage of 2015 (47.7%) and 2019 (78.4%) samples exceeded the 10 ppb (μg/kg) maximum level for total aflatoxins. For total fumonisins, the 2 ppm (mg/kg) maximum limit was not exceeded by any samples in 2015, but it was exceeded by 18.6% of 2019 samples [77].
Log10-transformed total aflatoxin (top) and total fumonisin (bottom) concentrations in all milled maize samples from the 2015 and 2019 Kongwa District surveys. Blue dashed lines indicate the maximum limits (ML) for total aflatoxins (10 ug/kg) and fumonisins (2 mg/kg) in maize flour. There were significantly higher concentrations of aflatoxins and fumonisins in 2019 than in 2015 (two-sample t-tests, **** p<0.0001).
The questionnaires were intended to interrogate various potential risk factors of mycotoxin contamination, and some survey responses were significantly associated with aflatoxin or fumonisin contamination. Flour type was associated with fumonisins in both 2015 and 2019; dehulled flours (sembe and kiwerege) had lower concentrations than whole grain flours (dona). This same reduction from dona to sembe was also observed for aflatoxin, but only among 2015 samples (Fig 6).
Divided by flour type, the log10-transformed total aflatoxin (top) and total fumonisin (bottom) concentrations in milled maize from the 2015 and 2019 Kongwa District surveys. Mycotoxin concentrations varied significantly across flour types (1-way ANOVA F<0.05) for 2015 aflatoxins, 2015 fumonisins, and 2019 fumonisins. Sembe (dehulled) flour had lower aflatoxins (2015) and fumonisins (2015 and 2019) than dona (whole grain), and kiwerege (dehulled and soaked) had lower fumonisins than dona in 2015 (two-sample t-test with Holm multiple testing adjustment, ns p>0.05, * p<0.05, ** p<0.01,*** p<0.001,**** p<0.0001,).
Modeling maize aflatoxins and fumonisins
Linear regression models from the 2015 survey offered key insights into the drivers of maize aflatoxin and fumonisin contamination. In modeling all samples, certain effects were shared between aflatoxin and fumonisin models; purchased maize increased aflatoxins and fumonisins, while kiwerege flour had a negative effect on both, and sembe flour was negatively associated only with fumonisins (Tables 2 and 3).
The aflatoxin model with only homegrown samples compared favorably to the aflatoxin model of only purchased samples, which was not significant. Also, the homegrown-only aflatoxin model explained more variation and identified more significant covariates than the full model: dehulled-soaked (kiwerege) flour, Soil PC1, and Soil PC2 (Table 2). This suggested that focusing only on homegrown maize could better model aflatoxins and identify more environmental risk factors.
When fumonisin data was divided into purchased and homegrown subsets, the homegrown model was significant (p-value = 0.027, but had no significant covariates, while the purchased model was not significant (p = 0.24) (Table 3). Together, the 2015 linear regression models showed that both survey and environmental factors explained variation in maize aflatoxins and fumonisins.
Nearly three times as many samples were used in the 2019 survey, and they were collected only at grain mills. Due to study design changes (e.g. a larger dataset with increased sampling per grain mill), the 2019 survey was better suited to the nested structure of a linear mixed model in which village was used as a random effect. The 2019 linear mixed model explained over a third of variation in aflatoxins, and it identified NIR PC1 and NDVI PC3 as significant covariates (Table 4). For maize fumonisins, a similar proportion of variation was explained using the same predictors, however, different significant covariates were identified (NIR PC3, NDVI PC1, NDVI PC2) (Table 4). In both models, a small amount of variation was explained by the random effects (mill village), suggesting that some local variation contributing to mycotoxin contamination was not captured by NIR, NDVI, or soil factors.
As in the 2015 survey, the 2019 models showed the influence of pre-harvest environmental factors on maize aflatoxin and fumonisin levels. NDVI covariates were more influential in 2019 models, perhaps because there was more variation in NDVI during this growing season. Furthermore, the 2019 models demonstrated the utility of including proximate sensing predictors, in the form of NIR reflectance measures, in complementing pre-harvest environmental factors.
Discussion
Understanding where, when, and to what extent mycotoxin contamination affects local communities is a necessary first step toward implementing scalable mitigation efforts. Over the course of two food surveys, we documented significant maize aflatoxin and fumonisin contamination within a Tanzania smallholder farming system. We sought to emulate the structure of a potential mycotoxin monitoring system by modeling these mycotoxins as a function of survey data, spectral grain analysis (NIR), and environmental factors.
Prior efforts to model mycotoxins with environmental data in SSA have sampled across agroecological zones and wide ranges of pre-harvest risk factors like soil quality, NDVI, rainfall, and temperature [46, 63, 78]. By contrast, this study entailed denser sampling within a semi-arid smallholder system. Even within a relatively narrow range of environmental variation, soil and NDVI factors explained variation in maize aflatoxins and fumonisins. Further, models for aflatoxins and fumonisins identified different significant environmental factors, which underscores the theory that although they frequently co-occur, pre-harvest aflatoxin and fumonisin contamination is differentially affected by the environment [21, 78]. In the future, the use of environmental data could be bolstered by higher temporal resolution datasets, mycotoxin-specific risk factors (e.g. drought stress index), and more precise knowledge of key events like planting date, flowering date, and harvest date.
We documented higher levels of maize aflatoxin and fumonisin contamination in the 2019 survey, and pre-harvest abiotic stress likely influenced this outcome. The plant health of millsheds in the 2018–2019 growing season, as measured by dekadal NDVI, was reduced compared to both the 2014–2015 season and to the historical norm. These reductions were especially prominent during key flowering, grain fill, and maturity periods. This reinforces the evidence that maize under abiotic stress during these critical periods is more susceptible to aflatoxin and fumonisin contamination [38, 39]. In 2019, farmer questionnaires also provided insights on environmental drivers of mycotoxin accumulation, as the majority reported that their matured maize was rained upon. Late season rains have been implicated as a pre-harvest risk factor for aflatoxins, as they promote A. flavus growth on already infected ears [38].
Our findings regarding maize processing in Kongwa District highlighted the interconnected tensions among food safety, food security, domestic labor, and cultural preferences. In Tanzania, dehulled maize flours (sembe or kiwerege) are often preferred as they produce whiter and softer ugali compared to whole grain flour (dona) [58]. Dehulled flours also take more time and labor, especially in the production of the highly valued kiwerege, which requires soaking dehulled maize in water for 1–6 days prior to milling [58]. From food security and nutritional perspectives, the dehulling process removes a significant fraction of edible maize (5–25% for sembe and 30–40% for kiwerege), and these dehulled flours are more carbohydrate-dense since they lack the nutrient-rich germ and bran [14]. Our findings indicated a preference for dehulled maize flours in Kongwa District, and confirmed the mycotoxin-reduction effect of dehulling, particularly for fumonisins [56, 58]. Processing techniques like dehulling should be accounted for, not only when monitoring maize mycotoxins, but also when considering them as part of mitigation efforts, as there are accompanying implications on food quality, food security, and local preferences.
Cost-effective and portable spectral devices, like the SCiO used in this study, have been tested for a variety of agricultural applications such as cultivar identification, cassava starch phenotyping, and quantifying milk nutrients [79–81]. We further demonstrated the utility of a relatively low-cost NIR spectrometer by using spectral information to both classify different maize flour types and model a portion of variation in maize aflatoxins and fumonisins. NIR signatures capture a wealth of information on the biochemical composition of maize grain, and they could help to model mycotoxins by indirectly accounting for associated risk factors (flour type, variety, soil quality). Future work could substantially improve on these efforts with improved spectrometers and larger sample sizes, and improved spectral detection of mycotoxins should be prioritized to expand mycotoxin monitoring, ideally validated by gold-standard quantification methods.
Homegrown maize has been cited as a prominent source of aflatoxins during notable aflatoxicosis outbreaks in eastern Kenya and in central Tanzania, close to where this survey occurred [11, 82, 83]. Although homegrown maize from the eastern Kenya hotspot had higher in aflatoxins than purchased maize, this study and others have documented the opposite, which suggests that this phenomenon is context-dependent [53, 83]. This is supported by observations in Kenya that maize farmers are more likely to sell lower quality grain with higher aflatoxin concentrations to markets and retain higher quality grain for home consumption [84]. Ultimately, the level of toxicity is likely to depend on a complex interaction of factors such as the source of the purchased maize, the toxicity of local fungal strains, the season during which samples are collected, and the storage quality and duration, among others [82, 85–87].
Ideally, mycotoxin monitoring could rapidly and efficiently identify hotspots and inform intervention efforts on how to respond in the most effective and appropriate manner; doing so would entail integrating community-level insights with centralized resources and capabilities in an appropriately flexible fashion [88]. Recent efforts to monitor other agricultural challenges in low-resource areas have deployed mobile apps that enable farmers to supply information and receive customized advice. PlantVillage’s Nuru app uses farmer-supplied images to identify the pests and diseases affecting their crops (e.g., viral diseases of cassava, fall armyworm, locusts), and the same images help improve prediction models based on remote-sensing environmental data [89]. Other lessons can be learned from plant disease surveillance efforts, such as MARPLE, which uses field-based genomics to rapidly identify fungal pathogen strains [90]. Mycotoxin monitoring schemes are difficult to implement due to the cost and infrastructure requirements for mycotoxin detection, but the frameworks and principles applied in Nuru and MARPLE could substantially inform these efforts.
This study aimed to serve as an example for how future monitoring efforts can better understand mycotoxin risk in smallholder agricultural contexts. We validated the importance of key elements of such a mycotoxin monitoring network: sampling at local nodes of maize processing, contextualizing the system with local feedback, capturing high resolution environmental variation, and using portable low-cost spectral sensors of grain quality. This work underscored the importance of using multiple indicators of mycotoxin risk in resource-limited and decentralized food systems. Scaling mycotoxin monitoring within smallholder settings is critical to predicting mycotoxin hotspots and ensuring a rapid response to a potential mycotoxicosis outbreak. It can also help to identify persistent factors that drive mycotoxin exposure and use them as a platform for sustainable mitigation efforts. This study demonstrated that integrating information on diverse risk factors bolstered mycotoxin modeling and improved understanding of mycotoxin dynamics.
Supporting information
S1 Appendix. 2019 grain mill survey questions in English and Swahili.
https://doi.org/10.1371/journal.pone.0316457.s001
(DOCX)
S1 Table. Compiled table of maize samples’ associated locations, mycotoxin concentrations, mycotoxin, soil, and NDVI values from the 2015 survey.
https://doi.org/10.1371/journal.pone.0316457.s002
(CSV)
S2 Table. Compiled table of maize samples’ associated locations, mycotoxin concentrations, soil, and NDVI values from the 2019 survey.
https://doi.org/10.1371/journal.pone.0316457.s003
(CSV)
Acknowledgments
The authors are thankful to Drs. Neema Kassim, Erica Phillips, Rebecca Stoltzfus, and the rest of the Mycotoxin Mitigation Trial team for advice and logistical support and for facilitating use of The Nelson Mandela African Institution of Science and Technology laboratory facilities. The authors are grateful for cooperation from the Kongwa District government, local community leaders, and survey participants. The authors would like to acknowledge Dr. Ying Sun for helpful feedback and guidance.
References
- 1.
IARC. Mycotoxin control in low- and middle-income countries. Wild CP, Miller DJ, Groopman JD, editors. Geneva, Switzerland; 2008. https://doi.org/10.2514/6.2008–1946
- 2. Mueller DS, Wise KA, Sisson AJ, Allen TW, Bergstrom GC, Bosley DB, et al. Corn yield loss estimates due to diseases in the United States and Ontario, Canada from 2012 to 2015. Plant Health Prog. 2016;17: 211–222.
- 3. Eskola M, Kos G, Elliott CT, Hajšlová J, Mayar S, Krska R. Worldwide contamination of food-crops with mycotoxins: Validity of the widely cited ‘FAO estimate’ of 25%. Crit Rev Food Sci Nutr. 2019; 1–17. pmid:31478403
- 4. CAST. Mycotoxins: Risks in plant, animal, and human systems. Task Force report, no. 139. Potential Economic Costs of Mycotoxins in the United States. 2003. Available: http://www.cast-science.org/publications/?mycotoxins_risks_in_plant_animal_and_human_systems&show=product&productID=2905
- 5. Ross RK, Yu MC, Henderson BE, Yuan JM, Qian GS, Tu JT, et al. Urinary aflatoxin biomarkers and risk of hepatocellular carcinoma. The Lancet. 1992;339: 943–946. pmid:1348796
- 6. Williams JH, Phillips TD, Jolly PE, Stiles JK, Jolly CM, Aggarwal D. Human aflatoxicosis in developing countries: A review of toxicology, exposure, potential health consequences, and interventions. Am J Clin Nutr. 2004;80: 1106–1122. pmid:15531656
- 7. Azziz-Baumgartner E, Lindblade K, Gieseker K, Rogers HS, Kieszak S, Njapau H, et al. Case-control study of an acute aflatoxicosis outbreak, Kenya, 2004. Environ Health Perspect. 2005;113: 1779–1783. pmid:16330363
- 8. Wild CP. Aflatoxin exposure in developing countries: The critical interface of agriculture and health. Food Nutr Bull. 2007;28. pmid:17658084
- 9. Shephard GS. Impact of mycotoxins on human health in developing countries. Food Addit Contam—Part Chem Anal Control Expo Risk Assess. 2008;25: 146–151. pmid:18286404
- 10. Kimanya ME, De Meulenaer B, Roberfroid D, Lachat C, Kolsteren P. Fumonisin exposure through maize in complementary foods is inversely associated with linear growth of infants in Tanzania. Mol Nutr Food Res. 2010;54: 1659–1667. pmid:20521269
- 11. Kamala A, Shirima C, Jani B, Bakari M, Sillo H, Rusibamayila N, et al. Outbreak of an acute aflatoxicosis in Tanzania during 2016. World Mycotoxin J. 2018;11: 311–320.
- 12. Turner PC, Collinson AC, Cheung YB, Gong Y, Hall AJ, Prentice AM, et al. Aflatoxin exposure in utero causes growth faltering in Gambian infants. Int J Epidemiol. 2007;36: 1119–1125. pmid:17576701
- 13. Cochrane N, D’Souza A. Measuring Access to Food in Tanzania: A Food Basket Approach. Econ Inf Bull. 2015 pp. 1–27.
- 14.
Ekpa O. Improvement of Maize-Based Foods in Sub-Saharan Africa. Wageningen University. 2020. Available: https://www.proquest.com/docview/2564080514?pq-origsite=gscholar&fromopenview=true&sourcetype=Dissertations & Theses
- 15. Mollay C, Kassim N, Stoltzfus R, Kimanya M. Complementary feeding in Kongwa, Tanzania: Findings to inform a mycotoxin mitigation trial. Matern Child Nutr. 2021;17: 1–10. pmid:33945210
- 16. Ngure FM, Kassim N, Phillips EL, Turner PC. Infant and Young Child Feeding Practices and Mycotoxins Contamination of Complementary Food Ingredients in Kongwa District, Tanzania. Curr Dev Nutr. 2023; 100030. pmid:37180082
- 17. Munkvold GP. Epidemiology of Fusarium Diseases and their Mycotoxins in Maize Ears. Eur J Plant Pathol. 2003;109: 705–713.
- 18. Cotty PJ, Mellon JE. Ecology of aflatoxin producing fungi and biocontrol of aflatoxin contamination. Mycotoxin Res. 2006;22: 110–117. pmid:23605583
- 19.
IARC Working Group on the Evaluation of Carcinogenic Risks to Humans. Some traditional herbal medicines, some mycotoxins, naphtalene, and styrene. IARC Monogr Eval Carcinog Risks Hum. Lyon, France: IARC; 2002.
- 20. Bandyopadhyay R, Kumar M, Leslie J. Relative severity of aflatoxin contamination of cereal crops in West Africa. Food Addit Contam. 2007;24: 1109–1114. pmid:17886182
- 21. Akello J, Ortega-Beltran A, Katati B, Atehnkeng J, Augusto J, Mwila CM, et al. Prevalence of aflatoxin-and fumonisin-producing fungi associated with cereal crops grown in zimbabwe and their associated risks in a climate change scenario. Foods. 2021;10: 1–18. pmid:33572636
- 22. Fandohan P, Zoumenou D, Hounhouigan DJ, Marasas WFO, Wingfield MJ, Hell K. Fate of aflatoxins and fumonisins during the processing of maize into food products in Benin. Int J Food Microbiol. 2005;98: 249–259. pmid:15698686
- 23. Kimanya ME, De Meulenaer B, Tiisekwa B, Ndomondo-Sigonda M, Devlieghere F, Van Camp J, et al. Co-occurrence of fumonisins with aflatoxins in home-stored maize for human consumption in rural villages of Tanzania. Food Addit Contam—Part Chem Anal Control Expo Risk Assess. 2008;25: 1353–1364. pmid:19680843
- 24. Ekpa O, Palacios-Rojas N, Kruseman G, Fogliano V, Linnemann AR. Sub-Saharan African Maize-Based Foods—Processing Practices, Challenges and Opportunities. Food Rev Int. 2019;35: 609–639.
- 25. Matumba L, Van Poucke C, Njumbe Ediage E, De Saeger S. Keeping mycotoxins away from the food: Does the existence of regulations have any impact in Africa? Crit Rev Food Sci Nutr. 2017;57: 1584–1592. pmid:25898143
- 26. Jayne TS, Mather D, Mghenyi E. Principal Challenges Confronting Smallholder Agriculture in Sub-Saharan Africa. World Dev. 2010;38: 1384–1398.
- 27. Morton JF. The impact of climate change on smallholder and subsistence agriculture. Proc Natl Acad Sci U S A. 2007;104: 19680–19685. pmid:18077400
- 28. Jones PG, Thornton PK. The potential impacts of climate change on maize production in Africa and Latin America in 2055. Glob Environ Change. 2003;13: 51–59.
- 29. Challinor A, Wheeler T, Garforth C, Craufurd P, Kassam A. Assessing the vulnerability of food crop systems in Africa to climate change. Clim Change. 2007;83: 381–399.
- 30. Ahmadalipour A, Moradkhani H, Castelletti A, Magliocca N. Future drought risk in Africa: Integrating vulnerability, climate change, and population growth. Sci Total Environ. 2019;662: 672–686. pmid:30703725
- 31. Warnatzsch EA, Reay DS, Camardo Leggieri M, Battilani P. Climate Change Impact on Aflatoxin Contamination Risk in Malawi’s Maize Crops. Front Sustain Food Syst. 2020;4: 1–13.
- 32. Alberts J, Rheeder J, Gelderblom W, Shephard G, Burger HM. Rural subsistence maize farming in South Africa: Risk assessment and intervention models for reduction of exposure to fumonisin mycotoxins. Toxins. 2019;11: 1–20. pmid:31212811
- 33. Kebede H, Abbas HK, Fisher DK, Bellaloui N. Relationship between aflatoxin contamination and physiological responses of corn plants under drought and heat stress. Toxins. 2012;4: 1385–1403. pmid:23202322
- 34. Vaughan MM, Huffaker A, Schmelz EA, Dafoe NJ, Christensen SA, McAuslane HJ, et al. Interactive effects of elevated [CO2] and drought on the maize phytochemical defense response against mycotoxigenic Fusarium verticillioides. PLoS ONE. 2016;11: 1–24. pmid:27410032
- 35. Battilani P, Camardo Leggieri M, Rossi V, Giorni P. AFLA-maize, a mechanistic model for Aspergillus flavus infection and aflatoxin B1 contamination in maize. Comput Electron Agric. 2013.
- 36. Chauhan Y, Tatnell J, Krosch S, Karanja J, Gnonlonfin B, Wanjuki I, et al. An improved simulation model to predict pre-harvest aflatoxin risk in maize. Field Crops Res. 2015;178: 91–99.
- 37. Janse van Rensburg B, McLaren NW, Flett BC. Grain colonization by fumonisin-producing Fusarium spp. and fumonisin synthesis in South African commercial maize in relation to prevailing weather conditions. Crop Prot. 2017;102: 129–136.
- 38. Cotty PJ, Jaime-Garcia R. Influences of climate on aflatoxin producing fungi and aflatoxin contamination. Int J Food Microbiol. 2007;119: 109–115. pmid:17881074
- 39. Cao A, Santiago R, Ramos AJ, Souto XC, Aguín O, Malvar RA, et al. Critical environmental and genotypic factors for Fusarium verticillioides infection, fungal growth and fumonisin contamination in maize grown in northwestern Spain. Int J Food Microbiol. 2014;177: 63–71. pmid:24607861
- 40.
Spoto F, Sy O, Laberinti P, Martimort P, Fernandez V, Colin O, et al. Overview of Sentinel-2. IEEE International Geoscience and Remote Sensing Symposium. Munich, Germany: Institute of Electrical and Electronics Engineers; 2012. pp. 1707–1710. https://doi.org/10.1109/IGARSS.2012.6351195
- 41. Hengl T, De Jesus JM, Heuvelink GBM, Gonzalez MR, Kilibarda M, Blagotić A, et al. SoilGrids250m: Global gridded soil information based on machine learning. PLoS ONE. 2017;12. pmid:28207752
- 42. McNally A, Arsenault K, Kumar S, Shukla S, Peterson P, Wang S, et al. A land data assimilation system for sub-Saharan Africa food and water security applications. Sci Data. 2017. pmid:28195575
- 43. Maxwell AE, Warner TA, Fang F. Implementation of machine-learning classification in remote sensing: An applied review. Int J Remote Sens. 2018;39: 2784–2817.
- 44. Kerry R, Ingram BR, Navarro F, Ortiz BV, Scully BT. Determining Corn Aflatoxin Risk within Counties in Southern Georgia, USA using Remotely Sensed Data. Adv Anim Biosci. 2017;8: 640–644.
- 45. Boken VK, Hoogenboom G, Williams JH, Diarra B, Dione S, Easson GL. Monitoring peanut contamination in Mali (Africa) using AVHRR satellite data and a crop simulation model. Int J Remote Sens. 2008;29: 117–129.
- 46. Smith L, Stasiewicz M, Hestrin R, Morales L, Mutiga S, Nelson R. Examining environmental drivers of spatial variability of aflatoxin accumulation in Kenyan maize: potential utility in risk prediction models. Afr J Food Agric Nutr Dev. 2016;3: 11086–11105.
- 47. Schaafsma AW, Hooker DC. Climatic models to predict occurrence of Fusarium toxins in wheat and maize. Int J Food Microbiol. 2007;119: 116–125. pmid:17900733
- 48. Damianidis D, Ortiz BV, Windham GL, Bowen KL, Hoogenboom G, Scully BT, et al. Evaluating a generic drought index as a predictive tool for aflatoxin contamination of corn: From plot to regional level. Crop Prot. 2018;113: 64–74.
- 49. Chauhan YS, Wright GC, Rachaputi NC. Modelling climatic risks of aflatoxin contamination in maize. Aust J Exp Agric. 2008;48: 358–366.
- 50.
Thomas TS, Robertson R, Boote K. Evaluating risk of aflatoxin field contamination from climate change using new modules inside DSSAT. International Food Policy Research Institute; 2019.
- 51.
Hell K, Fandohan P, Bandyopadhyay R, Kiewnick S, Sikora R, Cotty . Pre- and Postharvest Management of Aflatoxin in Maize: An African Perspective. In: Leslie JF, Bandyopadhyay R, Visconti A, editors. Mycotoxins: Detection methods, management, public health and agricultural trade. CABI; 2008. pp. 219–229. https://doi.org/dx.doi.org/10.1079/9781845930820.0309
- 52. Battilani P, Pietri A, Barbano C, Scandolara A, Bertuzzi T, Marocco A. Logistic regression modeling of cropping systems to predict fumonisin contamination in maize. J Agric Food Chem. 2008;56: 10433–10438. pmid:18841987
- 53. Mutiga SK, Hoffmann V, Harvey JW, Milgroom MG, Nelson RJ. Assessment of Aflatoxin and Fumonisin Contamination of Maize in Western Kenya. Phytopathology. 2015. pmid:25894319
- 54. Afolabi CG, Bandyopadhyay R, Leslie JF, Ekpo EJA. Effect of Sorting on Incidence and Occurrence of Fumonisins and Fusarium verticillioides on Maize from Nigeria. J Food Prot. 2006;69: 2019–2023. pmid:16924936
- 55. Mutiga SK, Were V, Hoffmann V, Harvey JW, Milgroom MG, Nelson RJ. Extent and drivers of mycotoxin contamination: inferences from a survey of Kenyan maize mills. Phytopathology. 2014;104: 1221–1231. pmid:24835222
- 56. Fandohan P, Ahouansou R, Houssou P, Hell K, Marasas WFO, Wingfield MJ. Impact of mechanical shelling and dehulling on Fusarium infection and fumonisin contamination in maize. Food Addit Contam. 2006;23: 415–421. pmid:16546888
- 57. Shu X, Livingston DP, Woloshuk CP, Payne GA. Comparative Histological and Transcriptional Analysis of Maize Kernels Infected with Aspergillus flavus and Fusarium verticillioides. Front Plant Sci. 2017;8: 1–14.
- 58. Ohna I, Kaarhus R, Kinabo J. No Meal without Ugali? Social Significance of Food and Consumption in a Tanzanian Village. Cult Agric Food Environ. 2012;34: 3–14.
- 59. Harvey J, Gnonlonfin B, Fletcher M, Fox G, Trowell S, Berna A, et al. Improving diagnostics for aflatoxin detection. In: Unnevehr L, Grace D, editors. Aflatoxins: Finding Solutions for Improved Food Safety. 2013.
- 60. Pearson TC, Wicklow DT, Pasikatan MC. Reduction of aflatoxin and fumonisin contamination in yellow corn by high-speed dual-wavelength sorting. Cereal Chem. 2004.
- 61. Stasiewicz MJ, Falade TDO, Mutuma M, Mutiga SK, Harvey JJW, Fox G, et al. Multi-spectral kernel sorting to reduce aflatoxins and fumonisins in Kenyan maize. Food Control. 2017;78: 203–214.
- 62. Aoun M, Siegel C, Windham GL, Williams WP, Nelson RJ. Application of reflectance spectroscopy to identify maize genotypes and aflatoxin levels in single kernels. World Mycotoxin J. 2022;15: 1–16.
- 63. Temba BA, Darnell RE, Gichangi A, Lwezaura D, Pardey PG, Harvey JJ, et al. The influence of weather on the occurrence of aflatoxin B1 in harvested maize from Kenya and Tanzania. Foods. 2021;10. pmid:33494323
- 64. Director KDE. Kongwa District: Social-Economic Profile. 2016.
- 65. Phillips E, Ngure F, Smith LE, Makule E, Turner PC, Nelson R, et al. Protocol for the trial to establish a causal linkage between mycotoxin exposure and child stunting: A cluster randomized trial. BMC Public Health. 2020;20: 1–11. pmid:32357944
- 66. Cooper D, Callahan B, Callahan P, Burnett L. Mobile Image Ratiometry: A New Method for Instantaneous Analysis of Rapid Test Strips. Nat Preced. 2012; 2–3.
- 67. Hollister J, Shah T, Robitaille AL, Beck MW, Johnson M. elevatr: Access Elevation Data from Various APIs. 2021.
- 68.
R Core Team. R: A Language and Environment for Statistical Computing. Vienna, Austria: R Foundation for Statistical Computing; 2022. Available: https://www.r-project.org/
- 69. Stevens A, Ramirez-Lopez L. An introduction to the prospectr package. 2022. Available: https://cran.r-project.org/web/packages/prospectr/vignettes/prospectr.html
- 70. Hijmans RJ. raster: Geographic Data Analysis and Modeling. 2023. Available: https://cran.r-project.org/package=raster
- 71. Bates D, Mächler M, Bolker B, Walker S. Fitting Linear Mixed-Effects Models Using {lme4}. J Stat Softw. 2015;67: 1–48.
- 72. Kuznetsova A, Brockhoff PB, Christensen RHB. {lmerTest} Package: Tests in Linear Mixed Effects Models. J Stat Softw. 2017;82: 1–26.
- 73. Bartoń K. MuMIn: Multi-Model Inference. 2022. Available: https://cran.r-project.org/package=MuMIn
- 74. Liaw A, Wiener M. Classification and Regression by randomForest. R News. 2002;2: 18–22.
- 75.
Wickham H. ggplot2: Elegant Graphics for Data Analysis. New York: Springer-Verlag; 2016. Available: https://ggplot2.tidyverse.org
- 76. Kassambara A. ggpubr: “ggplot2” Based Publication Ready Plots. 2023. Available: https://cran.r-project.org/package=ggpubr
- 77. EAC. East African Standard Maize Grains Specification. 2013 pp. 1–18.
- 78. Ng’ambi JT, Atehnkeng J, Monjerezi M, Ngongondo C, Vunain E, Ching’anda C, et al. Micro-climatic variations across Malawi have a greater influence on contamination of maize with aflatoxins than with fumonisins. Mycotoxin Res. 2022. pmid:36443622
- 79. Kosmowski F, Worku T. Evaluation of a miniaturized NIR spectrometer for cultivar identification: The case of barley, chickpea and sorghum in Ethiopia. PLoS ONE. 2018;13: 1–17. pmid:29561868
- 80. Riu J, Gorla G, Chakif D, Boqué R, Giussani B. Rapid analysis of milk using low-cost pocket-size NIR spectrometers and multivariate analysis. Foods. 2020;9. pmid:32785190
- 81. Hershberger J, Mbanjo EGN, Peteti P, Ikpan A, Ogunpaimo K, Nafiu K, et al. Low-cost, handheld near-infrared spectroscopy for root dry matter content prediction in cassava. Plant Phenome J. 2022;5.
- 82. Lewis L, Onsongo M, Njapau H, Schurz-Rogers H, Luber G, Kieszak S, et al. Aflatoxin contamination of commercial maize products during an outbreak of acute aflatoxicosis in eastern and central Kenya. Environ Health Perspect. 2005;113: 1763–1767. pmid:16330360
- 83. Daniel JH, Lewis LW, Redwood YA, Kieszak S, Breiman RF, Dana flanders W, et al. Comprehensive assessment of maize aflatoxin levels in eastern Kenya, 2005–2007. Environ Health Perspect. 2011;119: 1794–1799. pmid:21843999
- 84.
Hoffman V, Mutiga S, Harvey J, Nelson RJ, Milgroom M. Asymmetric information and food safety: Maize in Kenya. Agricultural & Applied Economics Association’s 2013 AAEA & CAES Joint Annual Meeting. Washington, DC; 2013. Available: https://ageconsearch.umn.edu/record/151288/
- 85. Hell K, Cardwell KF, Setamou M, Poehling HM. The influence of storage practices on aflatoxin contamination in maize in four agroecological zones of Benin, West Africa. J Stored Prod Res. 2000;36: 365–382. pmid:10880814
- 86. Probst C, Callicott KA, Cotty PJ. Deadly strains of Kenyan Aspergillus are distinct from other aflatoxin producers. Eur J Plant Pathol. 2012;132: 419–429.
- 87. Wenndt AJ, Sudini HK, Mehta R, Pingali P, Nelson R. Spatiotemporal assessment of post-harvest mycotoxin contamination in rural North Indian food systems. Food Control. 2021;126. pmid:34345120
- 88. Stafstrom W, Wenndt A, Nelson R. Mycotoxin Surveillance for Low-resource Settings. Mycotoxins in Food and Beverages Innovations and Advances Part I. 2021. pp. 1–29.
- 89. Mrisho LM, Mbilinyi NA, Ndalahwa M, Ramcharan AM, Kehs AK, McCloskey PC, et al. Accuracy of a Smartphone-Based Object Detection Model, PlantVillage Nuru, in Identifying the Foliar Symptoms of the Viral Diseases of Cassava–CMD and CBSD. Front Plant Sci. 2020;11: 1–14. pmid:33391304
- 90. Radhakrishnan GV, Cook NM, Bueno-Sancho V, Lewis CM, Persoons A, Mitiku AD, et al. MARPLE, a point-of-care, strain-level disease diagnostics and surveillance tool for complex fungal pathogens. BMC Biol. 2019. pmid:31405370