Oil Palm Yield Prediction Across Blocks Using Multi-Source Data and Machine Learning

Abstract

Predicting yields on a bigger scale in a timely and accurate manner is essential for preventing climate risk and ensuring food security, particularly in the light of climate change and the escalation of extreme climatic events. Furthermore, crop yield estimates are affected by various factors including weather, nutrients and management practices. In this study, integrating multi-source data (i.e. satellite-derived vegetation indices (VIs), satellite-derived climatic variables (i.e. land surface temperature (LST) and rainfall precipitation, weather station and field-surveys), we built one multiple linear regression (MLR), three machine learnings (XGBoost, support vector regression, and random forest) and one deep learning (deep neural network) model to predict oil palm yield at block-level within the oil palm plantation. Moreover, time-series moving average and backward elimination feature selection techniques were implemented at the pre-processing stage. The yield prediction models were developed and tested using MLR, XGBoost, support vector regression (SVR), random forest (RF) and deep neural network (DNN) algorithms. Their model performances were then compared using evaluation metrics and generated the final spatial prediction map based on the best performance. DNN achieved the best model performances for both selected (R2=0.91; RMSE= 2.92 tonnes per ha; MAE= 2.56 tonnes per ha and MAPE= 0.09) and full predictors (R2=0.76; RMSE of 3.03 tonnes per ha; MAE of 2.88 tonnes per ha; MAPE of 0.10 tonnes per ha). In addition, advanced ensemble ML techniques such as XGBoost may be utilised as a supplementary for oil palm yield prediction at the block level. Among them, MLR recorded the lowest performance. By using backward elimination to identify the most significant predictors, the performance of all models was improved by 5% - 26% for R2, and that decreased by 3% - 31% for RMSE, 7% - 34% for MAE, and 1% - 15% for MAPE, respectively. DNN generates the most accurate statistical metrics, with an increase of around 15% for R2, 11% for RMSE, 32% for MAE and 1% for MAPE. Our study successfully developed efficient, effective and accurate yield prediction models for timely predicting oil palm yield over a large area by integrating data from multiple sources. These can be potentially handled by plantation management to estimate oil palm yields to speed up the decision-making process for sustainable production.