Forecasting the Sugarcane Yields Based on Meteorological Data Through Ensemble Learning

Accurate prediction of sugarcane yields is crucial, particularly for developing countries like India, due to its economic significance and impact on farmers' livelihood. Unexpected fluctuations in production can affect farmers' income and the stability of the market, emphasizing the necessity of accurate forecasting to avoid adverse economic consequences. This research aims to enhance the precision of sugarcane yield prediction in India by developing a stacking ensemble learning model. The developed model incorporates the least absolute shrink and selection operator (LASSO), artificial neural network (ANN), and random forest (RF) as base models alongside random forest regression (RFR) and Ridge regression (RR) as meta-models and utilizes principal component analysis (PCA) and SHAPLEY values to reduce dimensions and explore feature correlations within the dataset. The data used in the study is obtained from ICRISAT and NASA databases covering 40 years (1982 to 2021) of meteorological information and sugarcane yield data across 24 districts of Uttar Pradesh, India. The model's generalizability is further improved through 5-fold cross-validation. For comparison, the vector autoregression moving average (VARMA) statistical method was also applied and it was observed that the outcome was not desirable. The findings indicate the competence of stacking ensemble model over individual models like LASSO, ANN, KNN, RF, and SVR.