Weather Pattern Analysis for Corn Yield Explainability

Weather is a principal driver of agricultural yield and exhibits both spatial and temporal heterogeneity with quasiperiodic behavior. This study integrates statistical summaries and topological features of weather time series to explain county-level corn yields in Indiana state of the US. Using persistent homology, we demonstrate that the geometric structure of weather data—capturing connected components and loops—varies across counties and years, and is highly sensitive to the temporal framing (e.g., full year vs. cropping season). Including daily extremes (maximum and minimum temperature) significantly enhances the topological signal compared to mean-based representations. Combining pattern descriptors–mean, standard deviation, entropy, low-frequency power, maximum jump–with topological invariants, i.e., connected components (H0 as distinct weather events) and loops (H1 as cyclical transitions in weather dynamics) counts, mean H1 persistence, and persistence gap, a machine learning model (CatBoost regression) achieved R2 = 0.59 and MAPE = 10.31% explaining 59% of the yield variability. Sensitivity analysis identified topological features—particularly the counts of H0, H1, and mean H1 persistence—as significant contributors to yield explanation, underscoring the value of topological structure in modeling agro-meteorological interactions.