An integrated model for pre- and post-harvest aflatoxin contamination in maize

The study addresses the growing risk of aflatoxin contamination in maize, a staple crop in many low- and middle-income countries where warm, humid climates and limited storage infrastructure facilitate Aspergillus flavus growth and aflatoxin production. The research question is how to predict and manage aflatoxin risk across the entire maize supply chain—from planting through storage to factory delivery—by integrating pre- and post-harvest dynamics driven by meteorology. Motivated by high shipment rejection rates at a large maize processing facility in Hyderabad, India, the authors aim to develop, parameterize, and validate a mechanistic model capable of predicting aflatoxin levels at the factory gate, informing sourcing decisions, and evaluating intervention strategies under dual cropping seasons and multiple sourcing regions.

Prior work has largely targeted pre-harvest risk, linking weather and crop stages to A. flavus infection and aflatoxin production. Mechanistic frameworks like the AFLA model (Battilani et al.) have been widely used and adapted to various crops and regions, while other approaches include empirical and crop growth–coupled models. However, it is well recognized that significant fungal growth and toxin accumulation continue post-harvest in storage, particularly under uncontrolled conditions prevalent in tropical and subtropical LMICs. There is a scarcity of granular data along the supply chain, especially for post-harvest phases. Recent literature highlights climate change as a driver of increased aflatoxin risk and the potential of predictive risk modeling for decision support. The integrated approach presented here extends earlier pre-harvest models by explicitly incorporating post-harvest processes, storage conditions, and sourcing logistics to address the full supply chain.

Data and setting: The study used 3-hourly meteorological fields (temperature, humidity, rainfall) from the UK Met Office Unified Model (10 km resolution), linearly interpolated to hourly, for 2011–2017 over a rectangular region covering three sourcing areas in India (Nizamabad, Guntur, Bellary) and the Hyderabad factory. Aflatoxin measurements (ppb, total aflatoxins) from daily shipments (2012–2017) to a MARS Inc. processing facility were used for model fitting (2012–2015) and validation (2016–2017). Sourcing shares by region and season (Kharif, Rabi) were available monthly; exact field origins were unknown. Model structure: A discrete-time, meteorology-driven compartmental model integrates pre-harvest, harvest-processing, and post-harvest storage dynamics with hourly time steps and spatially explicit meteorology across a 57×66 grid (~10×10 km cells). For each season, 1,000 simulated farms per region (3 regions × 2 seasons = 6,000 fields/year) were seeded with uniform sowing windows (Rabi: Oct 16–Nov 30; Kharif: Jun 1–Jul 31). Harvest occurs when a field accrues 1,500 growing degree days (GDD). After harvest processing, batches move to storage (30 days on farm, then markets), and daily shipments to the factory are sampled per historical sourcing proportions. Pre-harvest model: State variables (per unit area) updated hourly: N (soil spores), S (silk spores), F (kernel infection proportion/level), A (aflatoxin in kernels), and GDD. Core equations advance N, S, F, A, GDD driven by temperature and humidity, with susceptibility σ dependent on GDD. Key temperature/humidity-driven rates: sporulation α(T,H), liberation λ(T,H), deposition π(T,H), germination γ(T,H), infection and growth β_pre(T,H), aflatoxin production τ(T,H), thermal accumulation θ(T). Crops start free of spores, infection, and aflatoxin. Harvest processing: Includes optional filtering (ψ) that proportionally reduces F, A, and fines X, drying protection period δ during which growth and toxin production are halted, and bagging contamination χ_B adding fines X_B. For fitting, μ (fines proportionality), ψ (filtering efficacy), and χ_B were set to zero; δ was estimated. Post-harvest model: Tracks contaminants X (fines), infection F, and aflatoxin A in storage. Growth of A. flavus occurs from existing kernel infection and fines, with post-harvest growth rate β_post(T,H) and aflatoxin production τ(T,H) driven by ambient humidity and temperature (via water activity formulations). Some storage can be climate controlled by reducing T/H from ambient; otherwise ambient conditions prevail. Batches are stored on-farm (~30 days) then at markets until sourced or discarded after 1 year. Sourcing and sampling: Daily, shipments are randomly selected from available stored batches to match observed monthly region/season sourcing proportions. To mimic high within-batch sampling variance in observed data, model aflatoxin A is transformed to a comparable sampled value B ~ Exponential(mean = Ā) before comparison. Parameter estimation and evaluation: Five parameters were estimated using Approximate Bayesian Computation on 2012–2015 data: α (primary sporulation scaling), β_pre, β_post, τ_0 (aflatoxin production scaling, shared pre/post), and drying protection duration δ. A constrained uniform prior was sampled 750,000 times; the top 1% by a monthly 75th-percentile error metric were accepted to form the posterior. Model performance was assessed by monthly median aflatoxin levels and monthly rejection rates (>10 ppb threshold) vs. observed data, with accuracy bands: ±4 ppb for median aflatoxin and ±10% for rejection rates. Spatial maps of suitability (relative growth and toxin production rates) were generated from hourly meteorology.

• Parameter inference: ABC produced plausible posterior bounds for five key parameters. The drying protection period δ was tightly constrained to ~10–35 days and selected value δ_B ≈ 25 days. Selected parameter values (examples from Table 1): α_0 ≈ 1.0×10^-5, β_pre ≈ 6.31×10^-3, β_post ≈ 1.12×10^-3, τ_0 ≈ 5.06.
• Predictive performance: The integrated model replicated the timing and magnitude of seasonal peaks and monthly fluctuations in aflatoxin at the factory gate across fitting (2012–2015) and validation (2016–2017) periods. Monthly median aflatoxin accuracy: approximately 85% of months within ±4 ppb (Table 2: 83.7% fitting; 85.7% validation; 84.4% overall). The model tended to overpredict rather than underpredict (safer against false negatives).
• Rejection rates (>10 ppb): The model captured broad trends and peaks but overestimated during historically low-risk periods. Accuracy within ±10% of observed monthly rejection rates was ~50% (Table 3: 51.2% fitting; 52.4% validation; 51.6% overall), with a bias toward overprediction.
• Supply chain dynamics: Simulations show rapid pre-harvest colonization followed by substantial post-harvest growth and aflatoxin accumulation, highlighting storage as a major contributor to final aflatoxin levels. Pre-harvest status strongly conditions subsequent storage risk, especially for short storage durations.
• Spatial-temporal risk: Risk maps revealed strong spatial heterogeneity and temporal variability; monthly patterns can diverge from annual averages (e.g., September 2014 high inland risk vs. coastal annual average), emphasizing the need to consider timing and location jointly.
• Decision support and scenarios: Nowcasting example (Nizamabad Kharif 2016) indicated aflatoxin levels approaching the 10-ppb threshold and likely rises within two months, enabling proactive sourcing shifts. Scenario analyses showed: (i) filtration at harvest reduces aflatoxin proportionally to efficacy (strongest impact during high-risk periods), (ii) moderate cooling of market storage (5–10 °C below ambient) can significantly reduce aflatoxin with nonlinear benefits and dampened variability, and (iii) an optimized sourcing strategy (selecting lowest-risk batches monthly) substantially reduces peak aflatoxin levels compared with historical sourcing, providing a theoretical lower bound on achievable risk via sourcing decisions.

The integrated, meteorology-driven mechanistic model addresses the full maize supply chain, explicitly linking pre-harvest infection dynamics to post-harvest growth and toxin production under ambient or controlled storage. By fitting to real factory-gate time series via ABC, the model can reproduce observed aflatoxin patterns and provide practical decision support for sourcing and interventions. Overprediction tendencies make the tool conservative, reducing the frequency of missed high-risk periods (false negatives). Spatial and temporal variability in suitability underscores the importance of batch location histories and timing of movements. The model’s explicit biological processes, intervention levers, and sourcing logic enable sensitivity and scenario analyses valuable to industry, policy, and finance stakeholders. While rejection rate prediction was less precise than aflatoxin level prediction, the model nevertheless captured trends and extremes relevant to operational decision-making and risk mitigation.

The study develops and validates an integrated, mechanistic pre- and post-harvest model for A. flavus growth and aflatoxin production in maize, driven by high-resolution meteorology and informed by real sourcing and shipment testing data. It accurately reproduces monthly median aflatoxin levels and captures seasonal peaks while conservatively estimating rejection risk. The framework enables nowcasting, optimized sourcing, and evaluation of intervention strategies such as filtration and cooled storage to reduce contamination risk at the factory gate. Future work should focus on collecting measurements at intermediate supply chain points (A. flavus levels, aflatoxin, and moisture) to refine parameter estimates, explicitly modeling kernel moisture dynamics, incorporating additional environmental drivers (e.g., soil moisture), and coupling with economic analyses to identify cost-effective interventions and acceptable risk thresholds. The model is adaptable to other regions and is being trialed operationally to support sourcing and risk management decisions.

• Limited observational data along the supply chain: no direct measurements of A. flavus or aflatoxin during pre-harvest and storage; only factory-gate aflatoxin and regional sourcing shares were available for fitting/validation.
• Sampling variability: high within-batch variance required a sampling emulator, adding uncertainty to comparisons with historical data.
• Parameter identifiability: trade-offs among some parameters (e.g., α with β_pre and τ_0) due to data scarcity; posterior uncertainty persists despite ABC.
• Simplifications during fitting: filtering efficacy, fines contamination, and bag contamination were set to zero; drying represented as a fixed protection period rather than explicit moisture dynamics.
• Meteorological resolution and proxies: reliance on 10 km reanalysis and modeled water activity from ambient humidity; soil moisture and explicit kernel moisture dynamics not modeled.
• Rejection rate prediction less accurate than aflatoxin levels and biased toward overprediction, potentially affecting perceived regional risk without additional ground-truthing.
• Generalizability requires local calibration of sourcing patterns, storage practices, and intervention efficacy; practical feasibility of high filtration efficacy not established.