Segmentasi Semantik Multi-Kelas Citra Daun Kentang Berpenyakit Menggunakan Arsitektur U-Net++

Setyono, Emilia (2026) Segmentasi Semantik Multi-Kelas Citra Daun Kentang Berpenyakit Menggunakan Arsitektur U-Net++. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5002221028-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (21MB) | Request a copy

Abstract

Penyakit daun pada tanaman kentang (Solanum tuberosum L.) dapat menurunkan produktivitas tanaman apabila tidak diidentifikasi secara tepat. segmentasi semantik multi-kelas pada citra daun kentang berpenyakit, meskipun masih terdapat tantangan pada kelas dengan karakteristik visual yang saling menyerupaimulti-kelas pada citra daun berpenyakit memerlukan arsitektur yang mampu menangkap fitur dengan tingkat semantik yang konsisten antar level resolusi, sehingga U-Net++ dengan mekanisme Nested Skip Pathways dan Deep Supervision dipilih untuk mengatasi semantic gap yang menjadi keterbatasan arsitektur U-Net standar. Penelitian ini bertujuan untuk melakukan segmentasi semantik multi-kelas citra daun kentang berpenyakit menggunakan arsitektur U-Net++. Dataset yang digunakan diperoleh dari Roboflow sebanyak 2.815 citra dengan enam kelas penyakit, yaitu Bacteria, Fungi, Nematode, Pest, Phytophthora, dan Virus. Tahapan penelitian meliputi pra-pemrosesan data, pembagian data latih, validasi, dan uji, augmentasi selektif pada kelas minoritas, pelatihan model U-Net++, serta evaluasi menggunakan metrik Mean IoU, Dice Coefficient, dan Pixel Accuracy. Pelatihan model dilakukan menggunakan hybrid loss yang menggabungkan Cross-Entropy dan Dice Loss, serta pembobotan kelas berbasis IoU. Konfigurasi terbaik diperoleh melalui grid search terhadap parameter StepLR scheduler. Hasil pengujian menunjukkan bahwa model U-Net++ memperoleh Mean IoU sebesar 0,5630, Dice Coefficient sebesar 0,7192, dan Pixel Accuracy sebesar 0,8474. Hasil ini menunjukkan bahwa U-Net++ mampu melakukan segmentasi semantik multi-kelas pada citra daun kentang berpenyakit, meskipun masih terdapat tantangan pada kelas dengan karakteristik visual yang saling menyerupai.
=====================================================================================================================================
Leaf diseases in potato plants (Solanum tuberosum L.) can reduce crop productivity if they are not identified accurately. Image-based identification is one approach that can support disease observation in a more objective manner. Multi-class semantic segmentation of diseased leaf images requires an architecture capable of capturing features with a consistent level of semantic richness across resolution levels, motivating the use of U-Net++ with its Nested Skip Pathways and Deep Supervision mechanisms to address the semantic gap that limits the standard U-Net architecture. This study aims to perform multi-class semantic segmentation of diseased potato leaf images using the U-Net++ architecture. The dataset used in this study was obtained from Roboflow and consists of 2,815 images covering six disease classes, namely Bacteria, Fungi, Nematode, Pest, Phytophthora, and Virus. The research stages include data pre-processing, data splitting into training, validation, and test sets, selective augmentation on minority classes, U-Net++ model training, and evaluation using Mean IoU, Dice Coefficient, and Pixel Accuracy. Model training was performed using a hybrid loss that combines Cross-Entropy and Dice Loss, along with IoU-based class weighting. The best configuration was obtained through a grid search over the StepLR scheduler parameters. The testing results show that the U-Net++ model achieved a Mean IoU of 0.5630, a Dice Coefficient of 0.7192, and a Pixel Accuracy of 0.8474. These results indicate that U-Net++ is capable of performing multi-class semantic segmentation on diseased potato leaf images, although challenges remain for classes with visually similar characteristics.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Daun Kentang, Deep Learning, Penyakit Tanaman, Segmentasi Semantik, U-Net++, Potato Leaf, Deep Learning, Plant Disease, Semantic Segmentation, U- Net++
Subjects:	Q Science Q Science > QA Mathematics Q Science > QA Mathematics > QA336 Artificial Intelligence Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science)
Divisions:	Faculty of Science and Data Analytics (SCIENTICS) > Mathematics > 44201-(S1) Undergraduate Thesis
Depositing User:	Emilia Setyono
Date Deposited:	27 Jul 2026 02:36
Last Modified:	27 Jul 2026 02:36
URI:	http://repository.its.ac.id/id/eprint/137553

Actions (login required)

View Item