Oil palm yield prediction plays a vital role in supporting sustainable agricultural practices and guiding strategic decisions in the palm oil industry. With the increasing availability of historical and weatherrelated data, machine learning has become a promising approach for forecasting crop yields. This study evaluates the performance of both linear and non-linear machine learning models using historical agrometeorological data from 1986-2020 collected in Pahang, Malaysia. Specifically, we compare Linear Regression with three non-linear, tree-based models: Extra Trees, Random Forest and Gradient Boosting. The results show that the Extra Trees outperformed all other models explaining 88% of the variance (R²) in validation data with the lowest prediction error. Random Forest and Gradient Boosting also demonstrated strong performance with R² values of 79% and 78%, respectively. In contrast, Linear Regression achieved an R² of only 41%, indicating a limited ability to capture the non-linear relationships inherent in weather and environmental variables. This underperformance highlights the structural limitations of linear models when applied to complex agricultural datasets. Although non-linear models are computationally more demanding, their superior capacity to model complex, non-linear patterns makes them more suitable for real-world agricultural applications. These findings emphasise the value of tree-based machine learning models particularly Extra Trees in delivering reliable and accurate yield predictions, which are essential for sustainable oil palm plantation management.