Model Deteksi Pohon Kelapa Sawit dari Citra Drone dengan Pendekatan Deep Learning
Date
2022Author
Wibowo, Hery
Sitanggang, Imas Sukaesih
Mushthofa
Adrianto, Hari Agung
Metadata
Show full item recordAbstract
Indonesia merupakan produsen dan eksportir minyak sawit terbesar di dunia,
oleh karena itu pentingnya pengelolaan perkebunan kelapa sawit yang baik demi
keberlangsungan dan perkembangan komoditi kelapa sawit. Penghitungan pohon
kelapa sawit (tree counting) merupakan praktik perkebunan yang penting dilakukan
untuk inventarisasi aset biologis, estimasi produksi buah sawit, dan lain-lain.
Penerapan pertanian presisi dalam penghitungan pohon kelapa sawit dapat
diimplementasikan melalui deteksi pohon kelapa sawit dari citra udara.
Penelitian ini menggunakan pendekatan deep learning algoritme YOLOv3,
YOLOv4, dan YOLOv5m dalam mendeteksi pohon kelapa sawit dari citra drone.
Dataset diperoleh dari PT Perkebunan Nusantara VI berupa citra drone pekebunan
kelapa sawit di Jambi seluas 730 ha yang diberi 56.614 label kelas kelapa sawit.
Pengujian deteksi pohon kelapa sawit dilakukan pada citra drone seluas 180 ha.
Pohon kelapa sawit yang diteliti adalah tanaman muda (tahun tanam 2013–2015)
dan tanaman dewasa (tahun tanam 2009–2011) pada areal mendatar dan berbukit
dengan jarak antara kanopi tanaman jarang, berdekatan, dan tumpang tindih serta
pohon kelapa sawit berdampingan dengan vegetasi lain. Tujuan dari penelitian ini
membuat model menggunakan algoritme YOLOv3, YOLOv4, dan YOLOv5m
dalam mendeteksi pohon kelapa sawit dari citra drone. Evaluasi model berupa
perbandingan akurasi menggunakan confusion matrix dengan perhitungan Recall,
Precision, dan F1-score, serta waktu deteksi dan presisi bounding box. Tahapan
dari penelitian yaitu pengumpulan data, praproses data, pengembangan model,
pengujian model, dan evaluasi model. Hasil pengujian terhadap 17.343 pohon
kelapa sawit yang terbagi dalam 24 wilayah memperoleh F1-score untuk YOLOv3,
YOLOv4, dan YOLOv5m masing-masing sebesar 97,28%, 97,74%, dan 94,94%
dengan rata-rata waktu deteksi 43 detik, 45 detik, dan 21 detik untuk cakupan areal
seluas 12 ha per wilayah yang mencapai 1000 pohon. Hal ini menunjukkan akurasi
model sangat baik dan cepat dalam mendeteksi pohon kelapa sawit dan dapat
diterapkan pada perusahaan perkebunan.
Hasil dari penelitian ini diharapkan dapat membantu perusahaan perkebunan
dalam inventarisasi pohon kelapa sawit dengan akurasi tinggi dan efisien dalam
skala besar. Dengan pelaksanaan inventarisasi pohon yang tepat menjadi dasar
untuk pengelolaan perkebunan kelapa sawit yang lebih baik dan manajemen dapat
menjadikannya sebagai dasar penggunaan biaya. Indonesia is the largest producer and exporter of palm oil in the world,
therefore the importance of good oil palm plantation management for the
sustainability and development of the palm oil commodity. Tree counting is an
important plantation practice for biological asset inventories, fresh fruit bunch
production estimation, etc. The application of precision agriculture in counting oil
palm trees can be implemented by the detection of oil palm trees from aerial
imagery.
This study uses a deep learning approach using YOLOv3, YOLOv4, and
YOLOv5m in detecting oil palm trees from drone images. The dataset was obtained
from PT Perkebunan Nusantara VI in the form of drone images of oil palm
plantations in Jambi covering an area of 730 ha which were given 56,614 class
labels for oil palm. The detection test of oil palm trees was carried out on drone
images covering an area of 180 ha. The oil palm trees studied were young plants
(the planting year 2013–2015) and mature plants (the planting year 2009–2011) in
flat and hilly areas with sparse, close, and overlapping distances between plant
canopy, and oil palm trees intersect with other vegetations. The purpose of this
study is to create a model using the YOLOv3, YOLOv4, and YOLOv5m algorithms
in detecting oil palm trees from drone images. The evaluation of the model is in the
form of a comparison of accuracy using a confusion matrix with Recall, Precision,
and F1-score calculations, as well as detection time and bounding box accuracy.
The stages of the research are data collection, data preprocessing, model
development, model testing, and model evaluation. Model testing using images
from 24 regions, each of which covers 12 ha with up to 1000 trees (for a total of
17,343 oil palm trees) yielded F1-scores of 97.28%, 97.74%, and 94.94%, with an
average detection time of 43 seconds, 45 seconds, and 21 seconds for models
trained with YOLOv3, YOLOv4, and YOLOv5m, respectively. This result shows
that the method is sufficiently accurate and efficient in detecting oil palm trees and
has the potential to be implemented in commercial applications for plantation
companies.
The results of this study are expected to help plantation companies in
inventorying oil palm trees with high accuracy and efficiency on a large scale. With
proper implementation of tree inventories, it becomes the basis for better
management of oil palm plantations and management can use this as the basis for
using costs.