Short-Term Forecasting Arabica Coffee Cherry Yields by Seq2Seq over LSTM for Smallholder Farmers

Coffee production is a vital source of income for smallholder farmers in Mexico's Oaxaca, Chiapas, Puebla, and Veracruz regions. However, climate change, fluctuating yields, and the lack of decision-support tools challenge sustainable agricultural practices. The SABERES project aims to address these challenges through a Seq2Seq-LSTM model for predicting short-term coffee yields, utilizing datasets from Mexican national institutions, including the Agricultural Census (SIAP) and environmental data from CONAGUA. The model has demonstrated high accuracy in replicating historical yields for Chiapas and can forecast yields for the next two years. As a first step, we assessed coffee yield prediction for Beli, Indonesia, by comparing LSTM, ARIMA, and Step2Step-LSTM models using historical data. The results showed that Step2Step-LSTM provided the most accurate predictions, outperforming LSTM and ARIMA. Optimal performance was achieved using the maximum data sequence. Building on these findings, we aim to apply the best configuration to forecast coffee yields in Chiapas, Mexico. The Step2Step-LSTM model achieved an average difference of only 0.000247, with a similarity of 0.99975, indicating near-perfect accuracy. These results highlight the potential of Step2Step-LSTM in improving yield forecasts, supporting decision-making, and enhancing resilience in coffee production under climate change.