EMERGENCY OBJECT CLASSIFICATION IN AERIAL IMAGES USING RETINANET

Situmorang, Avia Maharani (2026) EMERGENCY OBJECT CLASSIFICATION IN AERIAL IMAGES USING RETINANET. Undergraduate thesis, Universitas Pembangunan Nasional “Veteran” Jawa Timur.

	Text (Cover) Thesis Cover.pdf Download (4MB)
	Text (Bab 1) Chapter 1-3-8.pdf Download (457kB)
	Text (Bab 2) Chapter 2.pdf Restricted to Repository staff only until 24 July 2029. Download (1MB)
	Text (Bab 3) Chapter 3.pdf Restricted to Repository staff only until 24 July 2029. Download (1MB)
	Text (Bab 4) Chapter 4.pdf Restricted to Repository staff only until 24 July 2029. Download (2MB)
	Text (Bab 5) Chapter 5.pdf Download (425kB)
	Text (Daftar Pustaka) References.pdf Download (293kB)

Abstract

Emergencies such as fires, floods, building collapses, and traffic accidents require rapid and accurate identification to support emergency response efforts. The use of aerial photographs can help obtain information about on-site conditions more comprehensively and efficiently, especially in areas that are difficult to reach directly. This study aims to analyze the performance of RetinaNet in classifying emergency objects in aerial photos using the Aerial Image Dataset for Emergency Response (AIDER). The research stages included data collection and annotation using LabelImg, data preprocessing, splitting the dataset into training, validation, and test sets, applying upsampling and downsampling techniques, training the RetinaNet model using the ResNet50 and ResNet101 backbones, and evaluating model performance using precision, recall, Average Precision (AP), and mean Average Precision (mAP). The dataset used consists of four emergency categories: collapsed buildings, fires, floods, and traffic accidents. Testing was conducted by varying the backbone, learning rate, batch size, and training data distribution. The results of the backbone testing on the original data show that ResNet101 achieved an mAP of 0.835790, which is higher than ResNet50, which achieved an mAP of 0.824750. Among all parameter testing scenarios, the best performance was achieved by RetinaNet with the ResNet101 backbone and a batch size of 8, with an mAP value of 0.841821. In the dataset variation tests, the combination of ResNet50 and upsampled data yielded the best performance with an mAP of 0.838928. Meanwhile, applying downsampling resulted in an mAP of 0.801831 for ResNet50 and 0.791151 for ResNet101. Upsampling techniques can improve the performance of ResNet50 but do not provide the same improvement for ResNet101, whereas downsampling tends to reduce performance due to a loss of training information. Based on these results, RetinaNet can be used to detect and classify emergency objects in aerial photos, with the optimal configuration being a ResNet101 backbone and a batch size of 8.

Item Type:

Thesis (Undergraduate)

Contributors:

Contribution	Contributors	NIDN/NIDK	Email
Thesis advisor	Anggraeny, Fetty Tri	NIDN0711028201	fettyanggraeny.if@upnjatim.ac.id
Thesis advisor	Junaidi, Achmad	NIDN0710117803	achmadjunaidi.if@upnjatim.ac.id

Subjects:

T Technology > T Technology (General)

Divisions:

Faculty of Computer Science > Departemen of Informatics

Depositing User:

AVIA MAHARANI SITUMORANG

Date Deposited:

24 Jul 2026 06:42

Last Modified:

24 Jul 2026 08:02

URI:

https://repository.upnjatim.ac.id/id/eprint/57636