Comparison of Support Vector Machine Optimization Methods for Chili Pepper Nutrient Deficiency Identification

Abstract

Deficiencies of macronutrients such as nitrogen, phosphorus, and potassium in chili pepper plants can cause visible changes in leaf characteristics, which can be utilized as information for identifying plant conditions through digital images. The performance of the Support Vector Machine in classifying leaf conditions is highly influenced by its hyperparameter configuration; therefore, an appropriate optimization strategy is required. This study aims to compare the performance of four Support Vector Machine optimization methods, namely Grid Search, Particle Swarm Optimization, Principal Component Analysis, and Adaptive Boosting, combined with RGB color features and Gray Level Co-occurrence Matrix texture features to classify four conditions of chili pepper leaves: healthy, nitrogen-deficient, phosphorus-deficient, and potassium-deficient leaves. The classification was performed using two image resolutions, namely 120 × 120 and 244 × 244 pixels. This study employed a quantitative approach using 1,040 chili pepper leaf images obtained from primary data and secondary datasets from Kaggle and Roboflow. The research process consisted of data augmentation, feature extraction, Support Vector Machine modeling, and evaluation using accuracy, precision, recall, and F1-score. The experimental results showed that the Particle Swarm Optimization configuration achieved the highest accuracy, reaching 86.06% for 244 × 244 pixel images and increasing to 91.83% for 120 × 120 pixel images. It was followed by Grid Search, which achieved 84.13% and 90.38%, respectively, while Principal Component Analysis and Adaptive Boosting achieved lower accuracy at both image resolutions. This study concludes that the combination of RGB and Gray Level Co-occurrence Matrix features with Particle Swarm Optimization provides the best classification performance for identifying nutrient conditions in chili pepper plants based on leaf images. The findings also indicate that image resolution affects classification accuracy differently across optimization methods.