Forecasting arabica coffee yields by auto-regressive integrated moving average and machine learning approaches

Abstract

Coffee is a major industrial crop that creates high economic value in Thailand and other countries worldwide. A lack of certainty in forecasting coffee production could lead to serious operation problems for business. Applying machine learning (ML) to coffee production is crucial since it can help in productivity prediction and increase prediction accuracy rate in response to customer demands. An ML technique of artificial neural network (ANN) model, and a statistical technique of autoregressive integrated moving average (ARIMA) model were adopted in this study to forecast arabica coffee yields. Six variable datasets were collected from 2004 to 2018, including cultivated areas, productivity zone, rainfalls, relative humidity and minimum and maximum temperatures, totaling 180 time-series data points. Their prediction performances were evaluated in terms of correlation coefficient (R²), and root means square error (RMSE). From this work, the ARIMA model was optimized using the fitting model of (p, d, q) amounted to 64 conditions through the Akaike information criteria arriving at (2, 1, 2). The ARIMA results showed that its R² and RMSE were 0.7041 and 0.1348, respectively. Moreover, the R² and RMSE of the ANN model were 0.9299 and 0.0642 by the Levenberg-Marquardt algorithm with TrainLM and LearnGDM training functions, two hidden layers and six processing elements. Both models were acceptable in forecasting the annual arabica coffee production, but the ANN model appeared to perform better.