Sistem Identifikasi Gulma Pada Padi Berbasis Deep Learning Menggunakan Metode You Only Look Once (YOLO)

Rachman, Risky Maulana (2024) Sistem Identifikasi Gulma Pada Padi Berbasis Deep Learning Menggunakan Metode You Only Look Once (YOLO). Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5009201118-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only until 1 October 2026. Download (5MB) | Request a copy

Abstract

Pada tanaman padi keberadaan gulma dapat menurunkan jumlah hasil panen padi, Dimana keberadaan gulma akan membuat tanaman padi bersaing untuk mendapatkan nutrisi, air, ruang, dan Cahaya, bahkan beberapa gulma dapat membawa penyakit bagi tanaman padi. Dari Latar belakang yang ada maka dapat dibuat sistem deteksi gulma pada tanaman padi. Untuk membuat sistem deteksi, dipilih model CNN dengan sistem One Stage Detector untuk sistem deteksi secara real-time, model yang dipilih YOLOv5. Dataset diperoleh dengan mengambil gambar padi dan gulma pada persawahan, didapatkan 650 gambar yang berisi padi dan dua jenis gulma yaitu Pistia Stratiotes, Ipomoea Aquatica, dan Monochoria Vaginalis yang sering tumbuh di persawahan Indonesia. Dataset akan diproses menjadi 2 jenis yaitu dataset Overlap dan Non-Overlap dan pembagian tiap data yaitu 80% Train dan 20% Test. Dengan Tuning Hyperparameter yang sama dilakukan pelatihan pada model, sehingga nantinya akan didapat 2 variasi model dan dibandingkan performansi dari tiap model, Dimana performansi yang dibandingkan adalah mAP, FPS, Number of Parameter, dan Training Time. Didapatkan data performansi dari model YOLOv5 untuk data Overlap dan Non-Overlap secara berturut-urut didapatkan mAP yaitu sebesar 88,68% dan 86,77%, dan FPS yaitu sebesar 24,45 fps dan 24,63 fps, Sedangkan untuk training time adalah sebesar 2.000 detik dan 1.800 detik, dan model YOLOv5 memiliki 7,02 juta parameter. Didapatkan Dari confusion matrix dari kedua jenis data, didapatkan bahwa model dapat mengenali objek crop, weed1 dan weed2 dengan baik yang dapat ditunjukkan pada jumlah deteksinya yang tinggi, walaupun memiliki jumlah yang tak terdeteksi atau background yang lumayan tinggi, sedangkan untuk weed3 sendiri memiliki jumlah deteksi yang rendah. Model YOLOv5 yang dirancang sudah cukup baik untuk deteksi objek yaitu gulma secara real-time.

========================================================================================================================

In rice plants the presence of weeds can reduce the amount of rice yields, where the presence of weeds will make rice plants compete for nutrients, water, space, and light, even some weeds can carry diseases for rice plants. From the existing background, a weed detection system can be made in rice plants. To create a detection system, a CNN model with a One Stage Detector system for real-time detection systems is chosen, the model chosen is YOLOv5. The dataset is obtained by taking pictures of rice and weeds in rice fields, obtained 650 images containing rice and two types of weeds namely Pistia Stratiotes, Ipomoea Aquatica, and Monochoria Vaginalis which often grow in Indonesian rice fields. The dataset will be processed into 2 types, namely Overlap and Non-Overlap datasets and the division of each data is 80% Train and 20% Test. With the same Hyperparameter Tuning, training is carried out on the model, so that later 2 variations of the model will be obtained and the performance of each model is compared, where the performance compared is mAP, FPS, Number of Parameters, and Training Time. Performance data from the YOLOv5 model for Overlap and Non-Overlap data is obtained, respectively, mAP is 88.68% and 86.77%, and FPS is 24.45 fps and 24.63 fps, While the training time is 2,000 seconds and 1,800 seconds, and the YOLOv5 model has 7.02 million parameters. From the confusion matrix of both types of data, it is found that the model can recognize crop, weed1 and weed2 objects well which can be shown in the high number of detections, although it has a fairly high number of undetected or background, while for weed3 itself has a low number of detections. The designed YOLOv5 model is good enough for real-time weed detection.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	CNN, Weed Detection, FPS, Paddy, YOLO, CNN, Deteksi Gulma, FPS, Padi, YOLO.
Subjects:	S Agriculture > SB Plant culture > SB191.R5 Rice farming T Technology > T Technology (General) > T11 Technical writing. Scientific Writing T Technology > T Technology (General) > T57.5 Data Processing T Technology > T Technology (General) > T57.62 Simulation T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK7882.P3 Pattern recognition systems
Divisions:	Faculty of Industrial Technology > Physics Engineering > 30201-(S1) Undergraduate Thesis
Depositing User:	Risky Maulana Rachman
Date Deposited:	12 Aug 2024 07:02
Last Modified:	12 Aug 2024 07:02
URI:	http://repository.its.ac.id/id/eprint/112526

Actions (login required)

View Item