Long-Term Dynamics and Forecasting in Türkiye’s Viticulture: A Comparative Analysis of Time Series and Machine Learning Models

Chemoecology · Springer Nature · 2026-11-06 · DOI: 10.1007/s10341-026-02080-z

Аннотация

Grapes are an economically important perennial fruit crop in Türkiye, yet long-term vineyard contraction and strong interannual production variability make forecasting challenging. This study evaluated Türkiye’s grape harvested area, production, and calculated yield (production/harvested area) during 1961–2024 using a strictly chronological validation framework. Naïve and drift benchmarks, linear regression, autoregressive integrated moving average (ARIMA), exponential-smoothing methods, decision tree, random forest, XGBoost, and support vector regression (SVR) were compared. Machine learning models used three autoregressive lags; SVR inputs and targets were standardized, and hyperparameters were selected using expanding-window forecasts during a separate 1990–1999 tuning period. Model robustness was assessed with 20 rolling-origin one-step forecasts for 2000–2019, while 2020–2024 was retained as an untouched five-step holdout. Accuracy was evaluated using root mean squared error (RMSE), mean absolute error (MAE), mean absolute scaled error (MASE), and symmetric mean absolute percentage error (sMAPE). Between 1961 and 2024, harvested area declined by 51.96%, production increased by 8.75%, and calculated yield rose by 126.39%. In rolling-origin validation, exponential smoothing had the lowest production RMSE (239,380 t), while Holt had the lowest yield RMSE (0.558 t/ha); several parsimonious models performed similarly for harvested area. In the final holdout, ARIMA achieved the lowest harvested-area RMSE (1387 ha), whereas decision tree had the lowest production RMSE (418,697 t) and yield RMSE (0.721 t/ha). Because production and yield rankings changed across evaluation windows, no single method was generally superior. The findings indicate that forecast performance depends strongly on both the target variable and the evaluation window.

Читать у издателя →

Похожие статьи