Oil Palm Yield Estimation Based on Vegetation and Humidity Indices Generated from Satellite Images and Machine Learning Techniques

Palm oil has become one of the most consumed vegetable oils in the world, and it is a key element in profitable global value chains. In Costa Rica, oil palm cultivation is one of the three crops with the largest occupied agricultural area. The objective of this study was to explain and predict yield in safe time lags for production management by using free-access satellite images. To this end, machine learning methods were performed to a 20-year data set of an oil palm plantation located in the Central Pacific Region of Costa Rica and the corresponding vegetation indices obtained from LANDSAT satellite images. Since the best correlations corresponded to a one-year time lag, the predictive models Random Forest (RF), Least Absolute Shrinkage and Selection Operator (LASSO), Extreme Gradient Boosting (XGBoost), Recursive Partitioning and Regression Trees (RPART), and Neural Network (NN) were built for a Time-lag 1. These models were applied to all genetic material and to the predominant variety (AVROS) separately. While NN showed the best performance for multispecies information (r(2) = 0.8139, NSE = 0.8131, RMSE = 0.3437, MAE = 0.2605), RF showed a better fit for AVROS (r(2) = 0.8214, NSE = 0.8020, RMSE = 0.3452, MAE = 0.2669). The most relevant vegetation indices (NDMI, MSI) are related to water in the plant. The study also determined that data distribution must be considered for the prediction and evaluation of the oil palm yield in the area under study. The estimation methods of this study provide information on the identification of important variables (NDMI) to characterize palm oil yield. Additionally, it generates a scenario with acceptable uncertainties on the yield forecast one year in advance. This information is of direct interest to the oil palm industry.