Implementation of long-term forecasting models in sugarcane for agricultural planning and yield goals setting

Abstract

This work aims to evaluate the efficiency of a long-term yield forecasting system for each minimum management unit (plot) at a sugar mill in Panama. Eleven teleconnections were used to forecast tons of cane per hectare (TCH) using Machine Learning (ML) with a lead time of 15-17 months before harvest. Three different ML models were trained, optimized, evaluated, and compared to select the best one for each plot. This process integrated individual plot-based predictions and measured overall efficiency. The Test results showed an average plot forecast of 79.00 TCH, while the average harvested TCH value was 78.59. Specifically, the plot-based models obtained an average Root Mean Square Error (RMSE) of 6.95 and a coefficient of determination (R2) of 0.8018. Notably, this system contributes to achieving TCH forecasts by plot with greater efficiency than conventional estimation by the farm manager. The 2024 harvest was influenced by warm El Niño-Southern Oscillation (ENSO) conditions and was used as a case study to exemplify the system’s applicability in managing work plans and adjusting them to the yield goals set by plot. To this end, individual plot yields were predicted in October 2022, resulting in a global average of 78.9 TCH. Based on the predicted results, a management plan was constructed and implemented in October 2022. This plan projected changes in management practices that would alter the expected production scenarios and unit costs. the cost per ton produced in the field decreased by 12.8 %, underscoring the effectiveness of the management strategies. The presented forecasting system could be implemented in other sugar mills, trained with local historical production data by plot, with the aim of being efficient in the use of inputs, minimizing environmental impact and taking actions that consider the effects of climate change on production