Comparison of MARS and Support Vector Regression Models for Estimating Pepper (Capsicum Annuum L.) Yield under Waterlogging Stress

Creative Commons License

Abdulrasaq S. A., Kiremit M. S., Arslan H.

Ondokuz Mayıs Üniversitesi Ziraat Fakültesi Dergisi (. Anadolu Tarım Bilimleri Dergisi), cilt.41, ss.599-622, 2026 (TRDizin)

Özet

Climate-driven increases in extreme rainfall events are exacerbating waterlogging stress, leading to significant yield losses and highlighting the need for a better understanding of plants’ adaptive responses. Accurately predicting yield responses to waterlogging is therefore critical for effective crop management. This study ai med to evaluate the effects of different waterlogging stages (vegetative, flowering, and fruiting) and durations (2, 4, 6, 8, and 10 days) on pepper yield, and to model these effects using Multivariate Adaptive Regression Splines (MARS) and Support Vector Regression (SVR). Both models successfully captured the relationship between waterlogging stress and yield; however, SVR demonstrated superior predictive performance. On the test dataset, SVR achieved a coefficient of determination (R²) of 0.946, root mean square error (RMSE) of 114.82 g, and mean absolute error (MAE) of 101.10 g, compared to R² = 0.876, RMSE = 177.09 g, and MAE = 142.72 g for MARS, representing reductions of 35.2% and 29.2% in RMSE and MAE, respectively. Variable importance analysis identified the waterlogging stage as the dominant factor influencing yield. The MARS model revealed a critical threshold of 6 days, beyond which yield declined sharply. SVR provides a more accurate and reliable approach for predicting pepper yield under waterlogging stress, while MARS offers greater interpretability by explicitly identifying threshold responses. These findings support the development of data-driven decision-support tools to manage waterlogging risk in pepper production systems.