ARDHI, Ovide Decroly Wisnu and Soeprobowati, Tri Retnaningsih and Adi, Kusworo and Prakasa, Esa (2026) IDENTIFIKASI PLANKTON PADA CITRA MIKROSKOPIK BERBASIS DEEP LEARNING. Doctoral thesis, UNIVERSITAS DIPONEGORO.

	Text cover aja.pdf Download (387kB)
	Text final_01_cover_sampai_sebelum_bab1.pdf Restricted to Repository staff only Download (2MB)
	Text 02_bab1.pdf Download (477kB)
	Text 03_bab2.pdf Download (4MB)
	Text 04_bab3.pdf Restricted to Repository staff only Download (2MB)
	Text 05_bab4.pdf Restricted to Repository staff only Download (46MB)
	Text 06_bab5.pdf Restricted to Repository staff only Download (476kB)
	Text 07_daftar_pustaka.pdf Download (502kB)
	Text 08_lampiran.pdf Restricted to Repository staff only Download (19MB)

Abstract

Plankton berperan sebagai produsen primer dan bioindikator kualitas perairan, namun identifikasi berbasis mikroskop masih memerlukan waktu yang panjang dan bergantung pada keahlian pakar, terutama pada spesies dengan morfologi halus dan tingkat kemiripan tinggi. Penelitian ini mengembangkan arsitektur deep learning ACViT–YOLO, yaitu modifikasi YOLOv8 yang mengintegrasikan Asymmetric Convolution (AC) untuk penguatan fitur lokal dan Vision Transformer (ViT) untuk pemodelan konteks global. Eksperimen dilakukan pada dataset cPID BRIN yang mencakup 20 spesies plankton dengan anotasi terbatas. Tahap pra-pemrosesan meliputi resizing, normalisasi, dan peningkatan kontras lokal menggunakan Contrast Limited Adaptive Histogram Equalization (CLAHE). Untuk mengatasi ketimpangan kelas yang dinyatakan sebagai Imbalance Ratio (IR), diterapkan augmentasi geometrik (rotasi, translasi, scaling, flipping, cropping) serta augmentasi fotometrik (penyesuaian kecerahan dan kontras, color jitter, dan penambahan noise). Evaluasi dilakukan menggunakan skema hold-out dan Stratified K-Fold Cross Validation (K = 5). Hasil menunjukkan bahwa pada kondisi ketimpangan tinggi, peningkatan F1 dan mAP berada pada kisaran 2–3%, dengan kenaikan precision mendekati 20%. Pada ketimpangan moderat, peningkatan agregat berada pada rentang 0,5–0,8%. Pada kondisi seimbang (IR 1), peningkatan mencapai 6–10% pada precision, recall, F1, dan mAP. Waktu inferensi meningkat 2–3% dibandingkan model dasar. Visualisasi EigenCAM menunjukkan perhatian model yang lebih terfokus pada struktur diagnostic objek. Temuan ini menegaskan bahwa ACViT–YOLO mampu menyeimbangkan akurasi deteksi, stabilitas, dan efisiensi komputasi, serta layak menjadi fondasi pengembangan sistem pendukung keputusan berbasis web yang lebih otomatis dan mudah diinterpretasikan untuk pemantauan plankton perairan tropis.
Kata kunci: plankton; deteksi objek; deep learning; ACViT–YOLO; YOLOv8; Vision Transformer; Imbalance Ratio.

Plankton function as primary producers and bioindicators of aquatic ecosystem quality; however, microscopy-based identification remains time-consuming and highly dependent on expert interpretation, particularly for species with subtle morphological characteristics and high inter-class similarity. This study develops a deep learning architecture termed ACViT–YOLO, a modification of YOLOv8 that integrates Asymmetric Convolution (AC) to enhance local feature representation and a Vision Transformer (ViT) for global contextual modeling. Experiments were conducted on the cPID BRIN dataset comprising 20 plankton species with limited annotations. Preprocessing stages included resizing, normalization, and local contrast enhancement using Contrast Limited Adaptive Histogram Equalization (CLAHE). To address class imbalance, quantified using the Imbalance Ratio (IR), geometric augmentation techniques (rotation, translation, scaling, flipping, and cropping) and photometric augmentation techniques (brightness and contrast adjustment, color jitter, and noise injection) were applied. Performance evaluation was conducted using a hold-out scheme and Stratified K-Fold Cross Validation (K = 5). Results indicate that under severe imbalance conditions, F1-score and mAP improvements range between 2–3%, with precision gains approaching 20%. Under moderate imbalance, aggregate improvements are limited to 0.5–0.8%. Under fully balanced conditions (IR = 1), performance gains reach 6–10% across precision, recall, F1-score, and mAP. Inference time increases by 2–3% relative to the baseline model. EigenCAM visualization reveals more focused attention on diagnostic object structures. These findings demonstrate that ACViT–YOLO effectively balances detection accuracy, performance stability, and computational efficiency, and provides a robust foundation for developing interpretable web-based decision support systems for tropical plankton monitoring.
Keywords: plankton; object detection; deep learning; ACViT–YOLO; YOLOv8; Vision Transformer; Imbalance Ratio

Actions (login required)

View Item