KURNIAWAN, Avip and Soeprobowati, Tri Retnaningsih and Warsito, Budi (2026) MODEL SISTEM INFORMASI PENYAKIT DAUN PADA TANAMAN PADI UNTUK KLASIFIKASI PENYAKIT BERBASIS PENGOLAHAN CITRA DAN MACHINE LEARNING. Doctoral thesis, UNIVERSITAS DIPONEGORO.

	Text cover aja.pdf Download (202kB)
	Text Cover lengkap.pdf Restricted to Repository staff only Download (1MB)
	Text BAB I.pdf Download (374kB)
	Text BAB II.pdf Download (731kB)
	Text BAB III.pdf Restricted to Repository staff only Download (1MB)
	Text BAB IV.pdf Restricted to Repository staff only Download (2MB)
	Text BaB V.pdf Restricted to Repository staff only Download (656kB)
	Text Daftar Pustaka.pdf Download (293kB)
	Text Lampiran.pdf Restricted to Repository staff only Download (518kB)

Abstract

Deteksi otomatis penyakit daun padi berbasis citra merupakan komponen penting dalam sistem pertanian cerdas, khususnya untuk membantu petani melakukan intervensi dini dan meminimalkan kerugian hasil panen. Penelitian ini mengevaluasi performa berbagai kombinasi tipe fitur warna, tekstur, bentuk, dan tepi serta metode seleksi fitur berbasis metaheuristik (Spyder Monkey Optimization/SMO, Particle Swarm Optimization/PSO, dan Ant Colony Optimization/ACO) untuk tugas klasifikasi tujuh jenis penyakit daun padi. Dua algoritma pembelajaran mesin utama digunakan, yaitu Random Forest (RF) dan Support Vector Machine (SVM) dengan kernel RBF, yang masing-masing diuji pada kondisi tanpa tuning dan dengan tuning hyperparameter. Hasil pengujian menunjukkan bahwa penggunaan fitur gabungan memberikan performa terbaik. Random Forest dengan fitur lengkap tanpa tuning mencapai akurasi tertinggi sebesar 94,35%, sedangkan SVM dengan fitur lengkap mencapai akurasi 93–94%. Sebaliknya, penggunaan fitur tunggal menghasilkan penurunan akurasi signifikan, terutama pada fitur bentuk yang hanya mencapai 32–55%, menandakan bahwa gejala penyakit daun padi bersifat kompleks dan tersebar pada berbagai komponen citra. Penerapan seleksi fitur menggunakan SMO, PSO, dan ACO mampu mempertahankan akurasi tinggi di kisaran 90–94%, namun tidak melampaui performa fitur gabungan. Temuan ini mengindikasikan bahwa eliminasi fitur tertentu dapat menghilangkan informasi diagnostik penting, sehingga seleksi fitur lebih cocok untuk kebutuhan efisiensi model dibandingkan peningkatan akurasi.Analisis per kelas menunjukkan bahwa kelas grassy stunt virus dan tungro virus lebih mudah dikenali model, sementara bacterial leaf blight, sheath blight, rice blast, dan healthy menjadi kelas yang paling menantang akibat kemiripan pola bercak dan variasi pencahayaan. Tuning hyperparameter memberikan pengaruh berbeda pada dua model: SVM mengalami peningkatan performa signifikan setelah penyetelan parameter C dan gamma, sedangkan RF justru mengalami penurunan akurasi ketika kedalaman pohon dan fitur per-node dibatasi. Secara keseluruhan, penelitian ini membuktikan bahwa kombinasi fitur penuh dan model berkapasitas tinggi khususnya Random Forest dan SVM-RBF tanpa reduksi fitur agresif merupakan pendekatan terbaik untuk klasifikasi penyakit daun padi pada dataset ini. Temuan ini dapat menjadi dasar pengembangan sistem deteksi tanaman berbasis citra yang akurat, ringan, dan berpotensi diimplementasikan pada aplikasi mobile di lapangan.
Kata Kunci : Klasifikasi penyakit daun padi, machine learning, Random Forest, Support Vector Machine, seleksi fitur.

Automated detection of rice leaf diseases using image analysis is an essential component of precision agriculture, enabling early intervention and reducing potential crop losses. This study evaluates the performance of multiple feature representations including color, texture, shape, and edge and three metaheuristic feature selection methods (Spyder Monkey Optimization/SMO, Particle Swarm Optimization/PSO, and Ant Colony Optimization/ACO) for classifying seven rice leaf disease classes. Two machine learning classifiers were employed, namely Random Forest (RF) and Support Vector Machine (SVM) with an RBF kernel, each tested under non-tuned and hyperparameter-tuned configurations.The experimental results demonstrate that feature fusion provides the best overall performance. Random Forest with full feature sets achieved the highest accuracy of 94.35%, while SVM with full features obtained 93–94%. In contrast, single-feature configurations resulted in significant performance degradation, particularly shape-based features, which only reached 32–55% accuracy. These findings indicate that disease symptoms are visually complex and distributed across multiple image characteristics. Although feature selection with SMO, PSO, and ACO retained relatively high accuracy (90–94%), none exceeded the performance of full feature fusion, suggesting that aggressive reduction may discard discriminative information and is therefore more suitable for model efficiency than peak accuracy.Class-wise evaluation shows that grassy stunt virus and tungro virus are consistently easier to distinguish, while bacterial leaf blight, sheath blight, rice blast, and healthy leaves remain challenging due to overlapping lesion patterns and illumination variations. Hyperparameter tuning produced different effects on the two models: SVM achieved substantial improvement after tuning C and gamma, whereas Random Forest experienced reduced performance when tree depth and per-node feature usage were restricted. Overall, the study confirms that combining all feature categories with high-capacity models particularly Random Forest and SVM-RBF without aggressive feature reduction provides the most accurate solution for rice leaf disease classification in this dataset. These findings offer practical guidance for developing lightweight, image-based disease diagnostic systems with potential deployment on mobile or edge devices in real farming environments.
Keywords: Rice leaf disease classification, machine learning, Random Forest, feature selection.

Actions (login required)

View Item