Implementasi Metode Hybrid CNN-KNN Berbasis Deep Learning untuk Klasifikasi Penyakit Tanaman Berdasarkan Citra Digital

Abstract

Chili plant diseases can reduce agricultural productivity, making fast and accurate identification important. This study develops a hybrid Convolutional Neural Network (CNN)-K-Nearest Neighbor (KNN) approach for chili leaf image classification. MobileNetV2 is employed as a feature extractor through transfer learning, followed by feature scaling and Principal Component Analysis (PCA) before KNN classification. KNN hyperparameters are optimized using GridSearchCV. The dataset contains eight classes—Chilli Anthracnose, Chilli Healthy, Chilli Leaf Curl Virus, Chilli Leaf Spot, Chilli Veinal Mottle Virus, Chilli Whitefly, Chilli Yellowish, and Non-Chilli—with 4,865 images. The data are divided into 3,332 training, 845 validation, and 688 testing images. Images are resized, normalized, and augmented for the training set. The best configuration uses K=3, distance weighting, cosine distance, and 200 PCA components. The final model achieves 94.43% accuracy, 94.64% precision, 94.43% recall, and 94.43% F1-score. The ablation study shows an improvement from 82.73% for the CNN baseline to 94.27% for CNN+KNN, 94.20% for CNN+KNN+GridSearchCV, and 94.43% for the complete CNN+PCA+KNN+GridSearchCV configuration. These results demonstrate that CNN-KNN hybridization improves classification performance, while PCA provides an additional improvement in the final configuration.