Comparative evaluation of machine learning algorithms for Maize Lethal Necrosis Disease detection under class imbalance.

Maize Lethal Necrosis Disease (MLND) remains a major threat to maize production in sub-Saharan Africa, creating a need for accurate, interpretable, and computationally efficient diagnostic tools. This study presents a comprehensive comparative evaluation of nine machine learning algorithms for MLND detection using a structured feature-based dataset comprising 7,892 maize leaf observations obtained from the Kenya Agricultural and Livestock Research Organization (KALRO). The extracted features captured color, texture, lesion, and morphological characteristics associated with MLND symptoms. To address moderate class imbalance, the Synthetic Minority Over-sampling Technique (SMOTE) and class weighting were incorporated within the model development pipeline, and their contribution was quantified through an ablation study. Model performance was evaluated using repeated stratified 5-fold cross-validation, an independent 80/20 test set, and leave-one-county-out cross-validation (LOCO-CV) to assess geographic generalisability. Across repeated cross-validation, the leading models demonstrated consistent predictive performance with narrow 95% confidence intervals. CatBoost achieved the highest mean accuracy (0.753; 95% CI: 0.730-0.773), while Logistic Regression attained the highest mean area under the receiver operating characteristic curve (AUC = 0.808; 95% CI: 0.787-0.830). The overlapping confidence intervals among the best-performing models indicate broadly comparable predictive performance, highlighting the importance of considering complementary criteria beyond accuracy alone. LOCO-CV demonstrated minimal performance variation across counties, indicating good internal geographic generalisability. Calibration analysis using Expected Calibration Error (ECE) showed that post-hoc probability calibration substantially improved the reliability of predicted probabilities. At the same time, explainable artificial intelligence techniques consistently identified Chlorosis Intensity, Green-Yellow Ratio, and Necrotic Area as the most influential predictors. Considering predictive performance alongside computational efficiency, model size, and inference latency, LightGBM provided the most favorable balance for resource-constrained environments. In contrast, Logistic Regression offered a highly interpretable alternative with minimal computational requirements. Although the achieved performance indicates that approximately one in four samples may be misclassified and further external validation is required before operational deployment, the findings demonstrate that structured feature-based machine learning provides a transparent, computationally efficient, and interpretable framework for MLND detection and establishes a strong foundation for future field-based decision-support systems.