Classification of oil palm female inflorescences anthesis stages using machine learning approaches

Prediction of pollination stages in oil palm plantations is an important research area in precision agriculture. Oil palm is known as the most efficient commercial crop with the capacity to fulfill the growing global demand for vegetable oil consumption. However, oil palm production dependence on pollination is experiencing issues with decreasing the actual yield. Consequently, alternative methods in commercial plantations such as human-assisted pollination and recently Wireless Sensor Network (WSN) have been deployed despite their high economic costs due to labor requirements. Oil palm assisted pollination requires precision, inspection, traceability, and validation processes in the field. Currently, all these processes are performed by humans that can be associated with false assumptions, uncertainty, and pollination latency. Therefore, Machine Learning (ML) approaches as a subset of the Artificial Intelligence (AI) domain provides efficient, cost-effective, and non-destructive solutions to determine these reproductive stages for future autonomous pollination system. Our goal was to reduce the variability of worker's performance in oil palms, using ML algorithms to make expert decisions and reduce the risk related to a transient workforce. This comparative empirical study examined and compared the performance of the Random Forest (RF) againstkNearest Neighbor (kNN) and Support Vector Machine (SVM) for classification of oil palm pre-anthesis and anthesis stages, dividing into four classes (1, 2, 3, and 4). These models were tested using thermal features (endogenous) individually and in combination with meteorological variables (exogenous). The performance of models is evaluated with specific measures of performance, such as overall user's and producer's accuracies and F-measure values derived from the confusion matrix. The results showed that the RF model produced better results with regard to average F-measure (88.6%, 71.83%), producer's accuracies (88.70%, 71.35%), and user's accuracies (88.27%, 72.36%) on test sets using exogenous + exogenous and endogenous feature sets, respectively.Among the three classifiers tested with two datasets, the RF method outperforms the other two popular algorithms, i.e.,kNN and SVM with respect to accuracy and F-measure metrics. The results validated the significance of thermal parameters, provided valuable features to devise an intelligent pollination management system, and proved the feasibility of using the RF model for the classification of oil palms four stages of anthesis.