Strategi Penanganan Ketidakseimbangan Data Citra Penyakit Daun Kelapa Sawit Menggunakan Class weight dan Generative Adversarial Network

Abstract

Ketidakseimbangan data masih menjadi tantangan dalam klasifikasi citra karena menyebabkan model cenderung mempelajari pola kelas mayoritas sehingga kemampuan mengenali kelas minoritas menurun. Berbagai strategi penanganan telah dikembangkan, di antaranya pendekatan berbasis algoritma melalui class weight dan pendekatan berbasis data melalui pembangkitan citra sintetis menggunakan Deep Convolutional Generative Adversarial Network (DCGAN). Penelitian ini bertujuan membandingkan efektivitas kedua strategi tersebut terhadap model baseline pada klasifikasi penyakit daun bibit kelapa sawit menggunakan MobileNetV2. Dataset yang digunakan terdiri atas 1.104 citra yang dikelompokkan ke dalam empat kelas, yaitu antraknosa, bercak daun, sehat, dan lainnya. Evaluasi dilakukan menggunakan Stratified 5-Fold Cross Validation dengan macro F1-score dan balanced accuracy sebagai metrik utama. Hasil penelitian menunjukkan bahwa skenario baseline menghasilkan rata-rata macro F1-score sebesar 89,07% dan balanced accuracy sebesar 87,00%, skenario class weight sebesar 90,05% dan 91,91%, sedangkan skenario DCGAN sebesar 91,19% dan 89,63%. DCGAN memberikan nilai macro F1-score tertinggi, sedangkan class weight menghasilkan balanced accuracy tertinggi. Namun, berdasarkan uji Kruskal–Wallis, tidak terdapat perbedaan performa yang signifikan antarketiga skenario (p-value > 0,05). Dengan demikian, penerapan class weight maupun DCGAN memberikan performa yang sebanding dalam menangani ketidakseimbangan data pada dataset yang digunakan.

Data imbalance remains a major challenge in image classification because it causes models to favor learning patterns from majority classes, thereby reducing their ability to recognize minority classes. Various strategies have been developed to address this issue, including algorithm-level approaches through class weighting and data-level approaches through synthetic image generation using Deep Convolutional Generative Adversarial Networks (DCGAN). This study aimed to compare the effectiveness of these strategies with a baseline model for oil palm seedling leaf disease classification using MobileNetV2. The dataset consisted of 1,104 images categorized into four classes: anthracnose, leaf spot, healthy, and others. Model performance was evaluated using Stratified 5-Fold Cross Validation, with macro F1-score and balanced accuracy as the primary evaluation metrics. The results showed that the baseline model achieved an average macro F1-score of 89.07% and a balanced accuracy of 87.00%, while the class weight approach achieved 90.05% and 91.91%, respectively. The DCGAN approach achieved a macro F1-score of 91.19% and a balanced accuracy of 89.63%. DCGAN produced the highest macro F1-score, whereas the class weight approach yielded the highest balanced accuracy. However, the Kruskal–Wallis test indicated no statistically significant difference in performance among the three scenarios (p-value > 0.05). Therefore, both class weighting and DCGAN provided comparable performance in handling data imbalance for the dataset used in this study.