Date of Defense

6-11-2025 5:00 PM

Location

H1-1116

Document Type

Thesis Defense

Degree Name

Master of Science in Remote Sensing and Geographic Information Systems

College

CHSS

Department

Geography and Urban Sustainability

First Advisor

Dr. Tareefa Alsumaiti

Keywords

Object Detection, Ship Detection, Satellite Imagery, Faster R-CNN, YOLOv3, SSD, Deep Learning

Abstract

The importance of rapid, reliable ship detection in satellite imagery is underscored by needs in maritime safety, environmental protection, and sustainable fisheries. In this study, a comparative assessment of three widely used object detectors—Faster R-CNN, YOLOv3, and SSD (300/512)—is presented to clarify how accuracy and speed are balanced for ship and dock detection. A unified pipeline (MMDetection) was employed so that model training, validation, and evaluation were standardized. ShipRSImageNet, a high-resolution dataset with COCO-style annotations, was used as the primary benchmark, while Airbus Ship Detection data were reformatted from masks to bounding boxes and standardized to COCO to ensure consistency. Performance was measured using COCO metrics (mAP@[0.50:0.95] with size-wise AP) and batch-1 latency to reflect realistic inference conditions.

In baseline experiments, superior accuracy was achieved by Faster R-CNN (mAP ≈ 55.1%), with particular gains observed on small vessels and cluttered scenes. Lower but competitive results were obtained by SSD512 (≈ 47.5%), YOLOv3-608 (≈ 45.5%), and SSD300 (≈ 42.1%). Speed advantages were exhibited by single-stage models: the highest throughput was recorded for SSD300 (≈ 73 images/s), followed by YOLOv3, SSD512, and then Faster R-CNN (≈ 32 images/s). Robustness was further examined under two constraints. When training was performed on down-scaled imagery and testing on full scale, smaller degradations were observed for Faster R-CNN and YOLO, whereas larger drops—most notably for SSD300—were recorded. Whenmodels were trained on only half of the data, greater resilience was demonstrated by YOLO,although the highest overall accuracy continued to be maintained by Faster R-CNN.

From these outcomes, guidance is provided for deployment: where fine-grained detection andsensitivity to small targets are prioritized, Faster R-CNN is recommended despite higherlatency; where near-real-time operation on constrained hardware is required, YOLO and SSDare indicated, with YOLO favored for data-limited settings and SSD512 for stronger single-stage precision. The results are anticipated to support model selection and scalableintegration in maritime monitoring systems.

Share

COinS
 
Nov 6th, 5:00 PM

A COMPARATIVE ANALYSIS OF YOLO, SSD, AND R-CNN MODELS FOR SHIP DETECTION IN SATELLITE IMAGERY

H1-1116

The importance of rapid, reliable ship detection in satellite imagery is underscored by needs in maritime safety, environmental protection, and sustainable fisheries. In this study, a comparative assessment of three widely used object detectors—Faster R-CNN, YOLOv3, and SSD (300/512)—is presented to clarify how accuracy and speed are balanced for ship and dock detection. A unified pipeline (MMDetection) was employed so that model training, validation, and evaluation were standardized. ShipRSImageNet, a high-resolution dataset with COCO-style annotations, was used as the primary benchmark, while Airbus Ship Detection data were reformatted from masks to bounding boxes and standardized to COCO to ensure consistency. Performance was measured using COCO metrics (mAP@[0.50:0.95] with size-wise AP) and batch-1 latency to reflect realistic inference conditions.

In baseline experiments, superior accuracy was achieved by Faster R-CNN (mAP ≈ 55.1%), with particular gains observed on small vessels and cluttered scenes. Lower but competitive results were obtained by SSD512 (≈ 47.5%), YOLOv3-608 (≈ 45.5%), and SSD300 (≈ 42.1%). Speed advantages were exhibited by single-stage models: the highest throughput was recorded for SSD300 (≈ 73 images/s), followed by YOLOv3, SSD512, and then Faster R-CNN (≈ 32 images/s). Robustness was further examined under two constraints. When training was performed on down-scaled imagery and testing on full scale, smaller degradations were observed for Faster R-CNN and YOLO, whereas larger drops—most notably for SSD300—were recorded. Whenmodels were trained on only half of the data, greater resilience was demonstrated by YOLO,although the highest overall accuracy continued to be maintained by Faster R-CNN.

From these outcomes, guidance is provided for deployment: where fine-grained detection andsensitivity to small targets are prioritized, Faster R-CNN is recommended despite higherlatency; where near-real-time operation on constrained hardware is required, YOLO and SSDare indicated, with YOLO favored for data-limited settings and SSD512 for stronger single-stage precision. The results are anticipated to support model selection and scalableintegration in maritime monitoring systems.