Tar Spot of Corn: Recent Insights and an Interpretable Weather–Imagery Pipeline for Disease Prediction

Tar spot of corn, caused by Phyllachora maydis (Maubl.), is characterized by black, tar-like lesions. The pathogen can reduce photosynthetic capacity, causing premature senescence and yield losses. After a decade in the U.S., P. maydis has resulted in an estimated loss of 937,000 bushels of corn in the U.S. and Canada. Now five years since the first review of tar spot, our understanding of this pathogen has substantially increased. Herein, the first section is a review of the recent advancements in knowledge of pathogen biology, inoculation protocols, modeling, and disease management, addressing ongoing challenges in tar spot dynamics. In the second section, an experimental pipeline comparing logistic regression and neural networks for interpretable tar spot prediction was explored. Current single-source models might fail to capture the full range of pathosystem variability, potentially leading to error biases and inaccurate predictions. Accordingly, we explored an innovative approach using multi-source data (weather and vegetation indices) to predict 1% tar spot severity at the ear leaf using neural networks (NN) and logistic regression (LR) models. Experiments conducted in Indiana (2021-2022) gathered 791 disease observations, 52 variables from multispectral images, and 90 weather station variables. Five modeling experiments were evaluated using different data sources, feature selection methods, and model performance. We used SHapley Additive exPlanations (SHAP) to enhance NN interpretability. Results showed NNs models outperformed LR in accuracy and specificity, particularly with vegetation indices (94.4% accuracy). LR models demonstrated higher sensitivity (96.1%) with weather data, suggesting strong prediction of tar spot at 1%. While data fusion showed inconsistent performance improvements, SHAP analysis and logistic regression pinpointed the minimum temperature during the night and relative humidity during the day, as well the standard deviation of Renormalized Difference Vegetation Index (RDVI) and Triangular Vegetation Index (TVI) as critical predictors. Future research should expand datasets across multiple locations, refine predictive thresholds, standardize disease assessment methods, and further integrate remote sensing technologies to enhance practical disease management applications to safeguard corn productivity.