A classification approach to wheat yield forecasting using machine learning and deep learning methods

Purpose. The purpose of the study is to model and classify wheat yield deviations in the Kherson region of Ukraine using machine learning and deep learning methods, taking into account both the internal dynamics of detrended yield residuals and external climate and economic predictors. Methodology. The study uses annual wheat yield and climate data for 1955–2021. Yield was detrended by piecewise linear trends reflecting technological and economic periods, and the residuals were transformed into a binary target: low yield and high yield. Two modelling strategies were compared: an endogenous approach based on lagged residuals and an exogenous approach using nine ten-day temperature indicators, monthly precipitation for April–June, and an economic-regime variable. Random Forest, Support Vector Machine, Logistic Regression, LSTM, and GRU models were evaluated using accuracy, precision, recall, F1-score, ROC-AUC, chronological train–test split, and 5-fold cross-validation. Results. The endogenous approach showed limited predictive ability: even LSTM and GRU did not exceed F1-score = 0.6667, which is explained by the short and weakly autocorrelated residual series. The exogenous approach substantially improved classification quality. In the chronological test sample, Logistic Regression achieved the highest F1-score (0.8571), while Random Forest identified all low-yield cases (recall = 1.0000). Cross-validation confirmed the stronger general discrimination of Random Forest (mean ROC-AUC = 0.81), whereas Logistic Regression remained valuable due to interpretability. Originality. The novelty lies in comparing endogenous and exogenous wheat-yield classification strategies using classical machine learning and recurrent neural networks. The study shows that deep learning does not provide a clear advantage when the yield-residual series is short and weakly autocorrelated, whereas climate and economic predictors markedly improve low-yield detection. Practical implications. The results can be used as a methodological basis for developing early-warning tools for low wheat yield risk under climate variability. The proposed modelling approach may support agricultural planning, risk assessment, and the selection of predictive models for regions where yield variability is influenced by both weather conditions and structural changes in the agricultural economy.