Perbandingan Performa CNN MobileNetV2 dan K-Nearest Neighbors untuk Klasifikasi Kondisi Tanaman Cabai

Abstract

Diseases and pests affecting chili plants can reduce crop quality and productivity if not detected at an early stage. The identification of plant conditions is still commonly performed through manual visual observation, making the process prone to diagnostic errors and time-consuming. This study aims to compare the performance of a MobileNetV2-based Convolutional Neural Network (CNN) and K-Nearest Neighbors (KNN) for chili plant condition classification using digital images. The dataset consisted of 7,991 training images, 515 validation images, and 355 testing images generated through a cropping process based on bounding-box annotations, covering seven categories of chili plant conditions: Anthracnose, Aphid, Armyworm, Healthy, Leaf Spot, Whitefly, and Yellowisha. Preprocessing included image resizing to 128 × 128 pixels and normalization. The CNN model employed a MobileNetV2 architecture with transfer learning, while the KNN model utilized manually extracted features consisting of Histogram of Oriented Gradients (HOG), Local Binary Pattern (LBP), and Color Histogram. Model performance was evaluated using a confusion matrix, accuracy, precision, recall, and F1-score with a weighted-average approach. The results show that CNN outperformed KNN, achieving an accuracy of 81.69%, precision of 86.28%, recall of 81.69%, and F1-score of 81.34%, whereas KNN achieved an accuracy of 45.63%, precision of 66.80%, recall of 45.63%, and F1-score of 42.24%. The contribution of this study lies in providing a comparative analysis between deep learning and traditional machine learning approaches on a dataset encompassing diseases, pests, and healthy conditions within the same classification scenario. The findings indicate that CNN's automatic feature extraction capability produces more effective visual representations than the manually engineered features used by KNN.