Pratama, Moh. Rizqi (2026) Analisis Kinerja Yolo Multi-Arsitektur Untuk Deteksi Dan Penghitungan Otomatis Tandan Buah Segar Kelapa Sawit Berbasis Drone Dengan Integrasi Dashboard Monitoring Web Real-Time. Masters thesis, Insistut Teknologi Sepuluh Nopember.
![[thumbnail of 6016251027- Master_Thesis.pdf]](http://repository.its.ac.id/style/images/fileicons/text.png) |
Text
6016251027- Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (6MB) | Request a copy |
Abstract
Taksasi produksi tandan buah segar (TBS) kelapa sawit secara manual bergantung pada penilaian subjektif mandor lapangan yang rentan terhadap inkonsistensi, tidak skalabel pada perkebunan luas, dan memerlukan waktu inspeksi yang panjang. Penelitian ini mengembangkan sistem Palm Eye, yaitu sistem deteksi, klasifikasi kematangan, dan penghitungan otomatis TBS berbasis drone yang terintegrasi dengan platform monitoring web real-time. Sistem menggunakan drone DJI Avata sebagai wahana pengambilan citra udara dan model YOLO12n sebagai detektor objek utama yang dipilih melalui perbandingan empiris terhadap enam generasi arsitektur YOLO (v5s, v8s, v9s, v10s, v11s, v12s) serta tiga ukuran model YOLO12 (nano, small, medium). Dataset terdiri dari 1.068 citra dengan empat kelas kematangan TBS, yaitu Ripe, Underripe, Unripe, dan Abnormal, yang dilatih menggunakan augmentasi berbobot kelas berbasis Albumentations untuk mengatasi ketidakseimbangan distribusi kelas dengan imbalance ratio awal sebesar 7,50. Model final YOLO12n dengan augmentasi menghasilkan mAP@0.5 sebesar 80,8%, mAP@0.5:95 sebesar 55,8%, precision 71,5%, dan recall 79,0% pada set uji independen. Penghitungan TBS diimplementasikan melalui integrasi algoritma ByteTrack dan mekanisme CountingSession dengan threshold stabilitas θ=60 frame, yang berhasil menurunkan mean absolute error (MAE) penghitungan pohon dari 47,3 (tanpa filter) menjadi 1,9. Pipeline penuh beroperasi pada kecepatan 35,6 FPS dengan latensi WebSocket 149 ms. Sistem dikembangkan menjadi platform web cloud-native berbasis React 18 dan FastAPI yang mencakup enam fitur: deteksi real-time, prakiraan panen berbasis model Accumulated Thermal Unit, peta satelit manajemen blok kebun, analisis NDVI menggunakan proxy VARI, riwayat sesi deteksi, serta estimasi volume dan bobot TBS menggunakan model elipsoid tri-aksial. Hasil penelitian menunjukkan bahwa integrasi drone, model YOLO generasi terbaru, dan platform web real-time mampu menghasilkan sistem taksasi TBS yang akurat, objektif, dan dapat dioperasikan langsung di lapangan perkebunan.
==================================================================================================================================
Manual yield assessment of oil palm fresh fruit bunches (FFB) relies on subjective field supervisor evaluations that are prone to inconsistency, unscalable across large plantations, and time-consuming. This study presents Palm Eye, an automated UAV-based system for FFB detection, maturity classification, and counting, integrated with a real-time web monitoring platform. The system employs a DJI Avata drone for aerial image acquisition and YOLO12n as the primary object detector, selected through empirical comparison of six YOLO architecture generations (v5s, v8s, v9s, v10s, v11s, v12s) and three YOLO12 model sizes (nano, small, medium). The dataset comprises 1,068 images across four FFB maturity classes Ripe, Underripe, Unripe, and Abnormal trained with class-weighted augmentation using Albumentations to address class imbalance (initial imbalance ratio of 7.50). The final YOLO12n model with augmentation achieved mAP@0.5 of 80.8%, mAP@0.5:95 of 55.8%, precision of 71.5%, and recall of 79.0% on an independent test set. FFB counting was implemented through the integration of ByteTrack and a CountingSession mechanism with a stability threshold of θ=60 frames, reducing tree counting mean absolute error (MAE) from 47.3 (without filtering) to 1.9. The full pipeline operates at 35.6 FPS with a WebSocket end-to-end latency of 149 ms. The system is deployed as a cloud-native web platform built on React 18 and FastAPI, encompassing six features: real-time detection, harvest forecasting via Accumulated Thermal Unit modeling, satellite block management mapping, NDVI analysis using VARI proxy, detection session history, and FFB volume and weight estimation using a tri-axial ellipsoid model. Results demonstrate that the integration of UAV technology, state-of-the-art YOLO architecture, and a real-time web platform yields an accurate, objective, and field-operable FFB yield assessment system.
Actions (login required)
 |
View Item |