Klasifikasi Family Pohon Berbasis Citra Uav Dengan Pendekatan Deep Learning

Mudhofar, Agus Fuad (2026) Klasifikasi Family Pohon Berbasis Citra Uav Dengan Pendekatan Deep Learning. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5024221021-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (13MB) | Request a copy

Abstract

Perubahan iklim global dan meningkatnya emisi gas rumah kaca mendorong kebutuhan akan sistem pemantauan hutan yang lebih akurat dan efisien. Indonesia, sebagai negara dengan kawasan hutan tropis terluas ketiga di dunia, memerlukan sistem inventarisasi vegetasi yang mampu mengidentifikasi komposisi pohon secara cepat untuk mendukung program konservasi, reforestasi, dan pengelolaan karbon. Metode konvensional yang mengandalkan pengamatan lapangan oleh tenaga ahli dinilai kurang skalabel untuk diterapkan pada wilayah yang sangat luas. Penelitian ini mengembangkan sistem deteksi dan klasifikasi famili pohon berbasis citra Unmanned Aerial Vehicle (UAV) menggunakan model deep learning YOLOv12. Fokus penelitian meliputi tiga famili pohon dominan di kawasan tropis Indonesia, yaitu Arecaceae (palem), Fabaceae (polong-polongan), dan Rubiaceae (kopi-kopian). Dataset terdiri atas 191 citra RGB beresolusi tinggi (4864 × 3648 piksel) yang diambil menggunakan drone DJI Phantom 4 Pro di Kebun Raya Purwodadi pada ketinggian 91,4 meter. Sebanyak 2.365 objek kanopi dianotasi secara manual, kemudian ditingkatkan menjadi 7.302 objek melalui augmentasi data menggunakan pustaka Albumentations. Pelatihan model dilakukan selama 300 epoch dengan resolusi masukan 640 × 640 piksel. Lima varian YOLOv12, yaitu Nano, Small, Medium, Large, dan Extra Large, dibandingkan untuk mengevaluasi keseimbangan antara akurasi dan efisiensi komputasi. Hasil penelitian menunjukkan bahwa varian YOLOv12l mencapai nilai mAP@0.5 tertinggi sebesar 0,7486 dengan precision 0,8046 dan recall 0,7006, sedangkan varian ringan YOLOv12n mampu berjalan pada perangkat Jetson Nano dengan kecepatan 8,62 FPS sehingga sesuai untuk pemantauan real-time di lapangan. Optimasi menggunakan TensorRT lebih lanjut meningkatkan kecepatan inferensi hingga 108,58 FPS pada GPU NVIDIA RTX 4090, sementara hasil kuantisasi menunjukkan bahwa perbedaan presisi (FP32, FP16, dan INT8) tidak memberikan pengaruh yang signifikan terhadap kecepatan inferensi pada perangkat uji. Perbandingan lintas arsitektur dengan Faster R-CNN dan EfficientDet menunjukkan keunggulan YOLOv12, di mana YOLOv12n memperoleh Macro F1-Score sebesar 0,6125, jauh melampaui Faster R-CNN (0,4572) dan EfficientDet (0,2609). Penelitian ini membuktikan bahwa kombinasi citra UAV dan deep learning dapat dimanfaatkan secara efektif untuk pemantauan keanekaragaman hayati, khususnya identifikasi famili pohon di kawasan hutan tropis, sekaligus membuka peluang penerapan pada perangkat komputasi terbatas di lapangan.
==================================================================================================================================
Global climate change and the increasing concentration of greenhouse gas emissions have created an urgent need for more accurate and efficient forest monitoring systems. Indonesia, home to the world’s third-largest tropical rainforest, requires vegetation inventory systems capable of rapidly identifying tree composition to support conservation, reforestation, and carbon management programs. Conventional methods based on field observations conducted by experts are considered insufficiently scalable for large forested areas. This study develops a tree family detection and classification system based on Unmanned Aerial Vehicle (UAV) imagery using the YOLOv12 deep learning model. The research focuses on three dominant tree families in Indonesia’s tropical forests: Arecaceae (palms), Fabaceae (legumes), and Rubiaceae (coffee family). The dataset consists of 191 high-resolution RGB images (4864 × 3648 pixels) captured using a DJI Phantom 4 Pro drone at Purwodadi Botanical Garden at an altitude of 91.4 meters. A total of 2,365 tree canopy objects were manually annotated and subsequently expanded to 7,302 objects through data augmentation using the Albumentations library. Model training was conducted for 300 epochs with an input resolution of 640 × 640 pixels. Five YOLOv12 variants—Nano, Small, Medium, Large, and Extra Large—were evaluated to assess the trade-off between accuracy and computational efficiency. The results show that the YOLOv12l variant achieved the highest mAP@0.5 of 0.7486, with a precision of 0.8046 and a recall of 0.7006, while the lightweight YOLOv12n variant achieved a processing speed of 8.62 FPS on a Jetson Nano device, making it suitable for real-time field monitoring. Further optimization using TensorRT increased the inference speed to 108.58 FPS on an NVIDIA RTX 4090 GPU, while quantization results demonstrated that differences in numerical precision (FP32, FP16, and INT8) had no significant impact on inference speed on the test platform. A cross-architecture comparison with Faster R-CNN and EfficientDet demonstrated the superiority of YOLOv12, with YOLOv12n achieving a Macro F1-Score of 0.6125, substantially outperforming Faster R-CNN (0.4572) and EfficientDet (0.2609). These findings demonstrate that the combination of UAV imagery and deep learning provides an effective solution for biodiversity monitoring, particularly for tree family identification in tropical forests, while also enabling deployment on resource-constrained edge devices in field environments.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	UAV, Citra Drone, Klasifikasi Famili Pohon, YOLOv12, Deep Learning, Deteksi Objek, Keanekaragaman Hayati, TensorRT, Jetson Nano, UAV, Drone Imagery, Tree Family Classification, YOLOv12, Deep Learning, Object Detection, Biodiversity, TensorRT, Jetson Nano
Subjects:	T Technology > T Technology (General) T Technology > T Technology (General) > T385 Visualization--Technique T Technology > T Technology (General) > T57.84 Heuristic algorithms.
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Computer Engineering > 90243-(S1) Undergraduate Thesis
Depositing User:	Agus Fuad Mudhofar
Date Deposited:	08 Jul 2026 09:00
Last Modified:	08 Jul 2026 09:00
URI:	http://repository.its.ac.id/id/eprint/134509

Actions (login required)

View Item