Deep Learning-Based Automatic Detection of Ships: An Experimental Study Using Satellite Images.

Area of Science:

• Remote Sensing
• Machine Learning
• Deep Learning

Background:

• Maritime surveillance is crucial for security and environmental protection.
• Automatic ship detection from satellite imagery is a key challenge with applications in traffic monitoring, illegal fishing prevention, and pollution control.
• Deep learning (DL) methods, particularly convolutional neural networks (CNNs), have shown significant promise in image recognition tasks.

Purpose of the Study:

• To develop and evaluate an automatic ship detection (ASD) approach using DL methods.
• To explore and compare different versions of the YOLO algorithm (YOLOv3, YOLOv4, YOLOv5) for ship detection in satellite images.
• To assess the performance of these algorithms on large satellite image datasets.

Main Methods:

• Utilized deep learning (DL) techniques for automatic ship detection (ASD).
• Compared YOLOv3, YOLOv4, and YOLOv5 algorithms for ship detection performance.
• Trained and evaluated algorithms on the Airbus Ship Challenge and Shipsnet datasets.

Main Results:

• All tested YOLO algorithms demonstrated effectiveness in detecting ships from satellite images.
• YOLOv5 achieved the highest accuracy at 99%.
• YOLOv4 and YOLOv3 achieved accuracies of 98% and 97%, respectively.

Conclusions:

• YOLOv5 is the superior algorithm for automatic ship detection among the evaluated YOLO versions.
• Deep learning approaches, specifically YOLOv5, offer a highly accurate solution for maritime surveillance using satellite imagery.