In-season Patterns and Environmental Drivers of Aflatoxin Contamination Across Peanut Seed Size Fractions

Abstract Aflatoxin contamination in peanut is highly variable across the growing season and may differ among seed fractions and locations within a field. Understanding its temporal dynamics, environmental associations, and spatial heterogeneity is essential for developing targeted aflatoxin risk-assessment and management strategies. This study aimed to characterize temporal, environmental, and spatial patterns of aflatoxin contamination in larger and smaller peanut seed fractions across commercial rainfed fields in Georgia and to evaluate the ability of environmental variables to predict contamination and deviations from expected seasonal patterns. Repeated in-season sampling was conducted in five commercial rainfed peanut fields. Sampling time was standardized using days after planting (DAP). Statistical inference, nonlinear seasonal modeling, and machine-learning analyses were integrated to evaluate seed-fraction differences, characterize seasonal contamination trajectories, and identify environmental predictors. Model performance was assessed using nested cross-validation with inner-fold hyperparameter optimization, permutation importance, TreeSHAP analysis, and leave-one-field-out validation. Spatial visualization was used to characterize within-field contamination heterogeneity across sampling events. Seed-fraction differences were seasonally dependent, with significantly greater contamination in the smaller fraction during middle sampling events but no significant differences at the beginning or end of the sampling period. Environmental prediction was stronger for direct contamination than for deviations from expected in-season trajectories. The best models explained 27.1% and 41.6% of direct-contamination variability in the larger and smaller fractions, respectively, compared with 10.9% and 18.6% of in-season-deviation variability. Permutation importance and TreeSHAP analyses consistently identified short-term moisture and thermal conditions, particularly recent rainfall, dry-day accumulation, and temperature metrics, as important predictive signals, although their relative contributions differed between seed fractions and prediction targets. Spatial visualization revealed substantial within-field heterogeneity across sampling events. Predictive performance declined when entire fields were withheld, indicating limited transferability to previously unseen fields. Overall, aflatoxin contamination exhibited temporally dynamic, spatially heterogeneous, and seed-fraction-specific patterns associated with short-term environmental conditions. These findings provide a framework for identifying periods and locations of elevated aflatoxin contamination risk and support the development of field-specific, temporally informed risk-assessment approaches. The limited transferability across fields highlights the need for broader multi-field validation before applying predictive models to new production environments or precision management decisions.