Classifying Nutrient Deficiencies in Palm Oil Leaves Using Convolutional Neural Network with Class Weights and Early Stopping Techniques

Abstract

Monitoring the health level of palm oil plants guarantees optimal production and quality oil yield. Experts have historically assessed plant health visually, but this method is limited by the number of samples that can be evaluated. Through automatic feature extraction, this study uses Convolutional Neural Networks (CNN) to classify nutrient deficiencies in oil palm leaves through leaf image analysis. Nevertheless, class imbalances in the dataset can lead to biased predictions. A CNN model with a class weight and an early stopping technique was developed to address this. A CNN model with three layers and a SoftMax activation function was trained over 200 epochs using Adam's optimizer and a categorical cross-entropy loss function to overcome this problem. According to the study, class weighting improves the classification accuracy of oil palm leaf photos. The classification accuracy of boron and potassium increased from 0.6119 to 0.7015 and 0.7500, respectively. However, magnesium classification still presents a challenge as accuracy drops to 0.4615, indicating the need for additional strategies to improve model performance across all nutrient classes.