Ramadhan, Muhammad Iqbal (2026) Sistem Terintegrasi Berbasis Deep Learning untuk Estimasi Berat Tandan Buah Segar Kelapa Sawit Secara Real-Time dari Live Stream Video Drone. Other thesis, Institut Teknologi Sepuluh Nopember.
![[thumbnail of 5025221274_Undergraduate_Thesis.pdf]](http://repository.its.ac.id/style/images/fileicons/text.png) |
Text
5025221274_Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (9MB) | Request a copy |
Abstract
Industri kelapa sawit memegang peranan vital dalam perekonomian Indonesia, namun proses estimasi hasil panen saat ini masih didominasi oleh metode manual. Drone komersial hanya dilengkapi kamera RGB monokular tanpa sensor kedalaman, sehingga informasi spasial tiga dimensi tidak tersedia secara langsung. Belum terdapat studi yang secara sistematis membandingkan kombinasi algoritma segmentasi instansi dan estimasi kedalaman monokular untuk mengestimasi berat TBS dari video drone. Penelitian ini mengembangkan sistem terintegrasi berbasis Deep Learning untuk mengestimasi berat TBS secara real-time dari live stream video drone komersial. Sistem mengatasi keterbatasan kedalaman melalui pendekatan konstruksi RGB-D (Red, Green, Blue, Depth), yaitu mengintegrasikan informasi warna dengan peta kedalaman yang diprediksi oleh model Monocular Depth Estimation (MDE). Metodologi dimulai dari ekstraksi frame video drone, dilanjutkan dengan instance segmentation untuk mengisolasi area TBS, estimasi kedalaman monokular untuk membentuk representasi RGB-D, rekonstruksi 3D berbasis point cloud yang diproyeksikan ke model elipsoid untuk kalkulasi volume, serta konversi volume ke berat menggunakan empat metode kalibrasi yang dibandingkan secara sistematis. Penelitian ini mengevaluasi empat kombinasi pipeline dengan memvariasikan segmentasi (YOLOv8-seg dan YOLOv8-det+SAM2) dan estimasi kedalaman (Depth Anything V2 dan DINOv2). Kombinasi YOLOv8-det+SAM2 dengan Depth Anything V2 menghasilkan akurasi estimasi berat terbaik dengan Mean Absolute Error (MAE) sebesar 8,14 kg dan Mean Absolute Percentage Error (MAPE) 35,7%. Temuan utama menunjukkan bahwa peningkatan kualitas segmentasi tidak berbanding lurus dengan peningkatan akurasi berat. Faktor pembatas utama adalah ambiguitas skala kedalaman yang inheren pada kamera monokular, bukan presisi segmentasi. Pipeline terbaik berhasil diimplementasikan sebagai sistem real-time yang memproses live stream video drone secara langsung di lapangan.
==================================================================================================================================
The palm oil industry plays a vital role in Indonesia’s economy, but the process of estimating crop yields is currently still dominated by manual methods. Commercial drones are equipped only with monocular RGB cameras without depth sensors, so three-dimensional spatial information is not directly available. No study has yet systematically compared the combination of object segmentation algorithms and monocular depth estimation to estimate the weight of fresh fruit bunches (FFB) from drone video. This research develops an integrated Deep Learning-based system to estimate the weight of FFB in real time from a live stream of commercial drone video. The system overcomes depth limitations through an RGB-D (Red, Green, Blue, Depth) approach, which integrates color information with the depth map predicted by the Monocular Depth Estimation (MDE) model. The methodology begins with the extraction of drone video frames, followed by instance segmentation to isolate TBS areas, monocular depth estimation to form an RGB-D representation, 3D reconstruction based on a point cloud projected onto an ellipsoid model for volume calculation, and the conversion of volume to weight using four calibration methods that are systematically compared. This study evaluates four pipeline combinations by varying the segmentation methods (YOLOv8-seg and YOLOv8-det+SAM2) and depth estimation methods (Depth Anything V2 and DINOv2). The combination of YOLOv8-det+SAM2 with Depth Anything V2 yielded the best weight estimation accuracy, with a Mean Absolute Error (MAE) of 8.14 kg and a Mean Absolute Percentage Error (MAPE) of 35.7%. Key findings indicate that improved segmentation quality does not directly correlate with improved weight accuracy. The primary limiting factor is the inherent depth scale ambiguity in monocular cameras, rather than segmentation precision. The best pipeline was successfully implemented as a real-time system that processes live drone video streams directly in the field.
| Item Type: | Thesis (Other) |
|---|---|
| Uncontrolled Keywords: | Deep Learning, Drone, Estimasi Berat, Monocular Depth Estimation, Point Cloud, Precision Agriculture, Tandan Buah Segar (TBS), Weight Estimation, Fresh Fruit Bunch (FFB) |
| Subjects: | Q Science > QA Mathematics > QA336 Artificial Intelligence Q Science > QA Mathematics > QA76.754 Software architecture. Computer software T Technology > T Technology (General) > T58.8 Productivity. Efficiency |
| Divisions: | Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis |
| Depositing User: | Muhammad Iqbal Ramadhan |
| Date Deposited: | 22 Jul 2026 09:03 |
| Last Modified: | 22 Jul 2026 09:03 |
| URI: | http://repository.its.ac.id/id/eprint/136291 |
Actions (login required)
 |
View Item |