Oil Palm Yield Gap Prediction Using Machine Learning: A Proof of Concepts

Abstract

The yield gap - the difference of potential yield and actual farm yield is a major indicator of yield-limiting factors. Recent advancements in precision agriculture have highlighted the importance of innovative methods for crop yield gap estimation. Weather variations influencing oil palm yield emphasized the need for exploring underlying yield gap patterns for timely forecast and policymaking to minimize it. This work proposed a supervised machine learning approach for the prediction of the oil palm yield gap in Pahang state Malaysia under fluctuating environmental conditions. Multisource data and machine learning methods combined with domain knowledge were explored. The procedure was carried out by (1) exploratory data analysis, (2) data preprocessing, (3) feature selection, (4) model selection, (5) model training, (6) hyper parameter tuning, and (7) performance evaluation. Additionally, features interdependency and the correlations with yield gap were also examined. The performance of the random forest regression model was assessed using six evaluation metrics including mean squared error and coefficient of determination R2 as the key performance indicator. Results suggested that wind speed, maximum temperature, and surface pressure have a positive association while optimum rainfall, root zone soil moisture, and solar radiation are negatively correlated with the yield gap. The R2 of 0.933 for train and 0.798 for test data with low error values indicated good prediction accuracy. Moreover, statistical measures such as model-based feature importance, learning curve, k-fold, validation curve, error, and residual analysis helped to investigate the model's learning process. The results prove strong potential of machine learning models to analyze, predict and close oil palm yield gaps.