VISVESVARAYA TECHNOLOGICAL UNIVERSITY

Jnana Sangama, Belagavi590018

A PROJECT SYNOPSIS
ON
“Ship Detection from Satellite images using Deep Learning”
Submitted in partial fulfilment of the requirement for award of degree
of
BACHELOR OF ENGINEERING
in
ARTIFICIAL INTELLIGENCE AND DATA SCIENCE
by

Vishal MP 1EP21AD059
Vishnu G 1EP21AD060

Under the guidance of

Dr. Anand R
HOD
Dept. of AI & DS, EPCET

20242025
Ship Detection from Satellite images using Deep Learning

CONTENTS

1. Abstract …………………………………………………………………………………. 03
2. Driven Force by Idea …………………………………………………………………… 04
3. Literature Review ………………………………………………………………………. 05
3.1. Paper 1 ……………………………………………………………………………… 05
3.2. Paper 2 ……………………………………………………………………………… 05
3.3. Conclusion ………………………………………………………………………….. 05
4. Problem Statement and Objectives …………………………………………………….. 06
4.1. Problem Statement ………………………………………………………………….. 06
4.2. Objectives …………………………………………………………………………... 06
5. System Design …………………………………………………………………………… 07
5.1. Proposed Architecture ……………………………………………………………… 07
5.2. CNN Architecture …………...……………………………………………………… 08
5.3. Hardware and Software Requirements ……………………………………………... 08
5.4. Tools and Libraries ………………………………………………………………… 08
6. Conclusion ……………………………………………………………………………….. 09
7. References…….……………………………………………………………………………..10

Dept. of AI&DS2024 - 2025 Page | 2

Ship Detection from Satellite images using Deep Learning

CHAPTER 1: ABSTRACT

Detecting ships in satellite imagery plays a crucial role in maritime security, traffic monitoring,
and environmental protection. However, traditional methods often struggle due to the vast
ocean background, varying weather conditions, and the small size of ships in high-resolution
satellite images. These challenges demand the development of an advanced and automated
solution to improve accuracy and reliability in ship detection tasks.

This project aims to leverage deep learning techniques to build an efficient system for ship
detection in satellite imagery. A convolutional neural network (CNN)based model, such as
YOLO or Faster RCNN, is designed to accurately identify ships amidst complex oceanic
environments. The system is trained and validated using diverse datasets to address challenges
like low visibility, varying lighting conditions, and overlapping objects, ensuring robustness
and adaptability in realworld scenarios.

The proposed system demonstrates significant improvements in detection accuracy and

reliability compared to traditional methods. It has practical applications in maritime security,
illegal activity detection, traffic monitoring, and environmental protection. By automating ship
detection with deep learning, this project provides a scalable and efficient solution to address
pressing challenges in the maritime domain.

Guide Remarks:

Signature with date:

Project Coordinator Signature

Dept. of AI&DS2024 - 2025 Page | 3

Ship Detection from Satellite images using Deep Learning

CHAPTER 2: DRIVEN FORCE BEHIND THE IDEA

The increasing reliance on maritime activities for global trade, transportation, and resource
exploration has underscored the need for efficient ship detection systems. However, manual
methods and traditional approaches often fall short when analyzing largescale satellite imagery
due to vast ocean backgrounds, unpredictable weather conditions, and the tiny size of ships in
high-resolution images. The growing demand for realtime monitoring and maritime security
has been a driving force in seeking automated solutions to address these limitations.

Advancements in deep learning have revolutionized computer vision tasks, including object
detection. The idea of utilizing stateoftheart models like YOLO and Faster RCNN for ship
detection stems from their proven capabilities in handling small object detection and complex
environments. These models, coupled with highquality satellite imagery, provide an
opportunity to bridge the gap between traditional limitations and the accuracy required for
maritime applications.

Another significant motivator is the potential impact on environmental and security challenges.
Automated ship detection can help in identifying illegal fishing activities, monitoring oil spills,
and tracking unauthorized or suspicious vessels. Such capabilities are crucial for preserving
marine ecosystems and ensuring maritime safety, making the development of an advanced ship
detection system not only a technical challenge but also a pressing societal need.

Furthermore, the availability of opensource datasets and accessible deep learning frameworks
has democratized the ability to tackle this problem. The driving force behind this project is to
combine these resources with innovative AI techniques to create a scalable, robust, and realtime
solution that can support industries, governments, and environmental organizations in
achieving their maritime goals. This project embodies the convergence of technological
innovation and realworld problemsolving for a safer and more sustainable maritime future.

Dept. of AI&DS2024 - 2025 Page | 4

Ship Detection from Satellite images using Deep Learning

CHAPTER 3: LITRATURE REVIEW

3.1. Vehicle Detection in Satellite Video Using a Compact 3×3 CNN and NMS
Authors: Roman Pflugfelder, Axel Weissenfeld, Julian Wagner (2022)

This paper presents a vehicle detection method using a compact 3×3 Convolutional Neural
Network (CNN) in conjunction with Non-Maximum Suppression (NMS) for satellite video
analysis. The proposed model is efficient, delivering high F1 scores of 0.81 and 0.87 despite
minimal labeled data. This makes the method highly resource-friendly and suitable for
applications where labeled data is scarce or costly. By utilizing a compact CNN architecture,
the model is optimized for satellite video analysis, offering a balance between performance and
computational efficiency.

The main advantages of this approach are its efficiency in satellite video applications, minimal
data requirements for training, and high detection accuracy. However, a notable limitation is
the model's generalizability, as it may not perform as well on different datasets beyond the ones
it was originally trained on.

3.2. Active Learning for Efficient Object Detection

Authors: Kazuki Uehara and Co. (2019)

In this paper, the authors introduce an active learning strategy designed to significantly reduce
the need for labeled samples in object detection tasks. By utilizing this approach, the amount
of labeled data required is reduced by 95%, while still maintaining a high detection accuracy
of 94%. This method is shown to outperform traditional object detection methods by 30%,
providing a substantial improvement in efficiency while ensuring high-quality results.

The key strengths of this approach include its drastic reduction in labeled sample requirements
and its ability to maintain a high level of detection accuracy. Additionally, it outperforms
conventional methods by a considerable margin. However, one potential downside is that this
strategy may not be applicable to all types of object detection tasks, suggesting that its
effectiveness might vary depending on the specific context or dataset used.

3.3. Conclusion

Both papers propose efficient object detection methods with minimal labeled data. Paper 1 uses
a compact CNN and NMS for vehicle detection in satellite videos, achieving high accuracy but
with limited generalizability. Paper 2 employs an active learning strategy that reduces labeled
samples by 95%, maintaining 94% accuracy, though its applicability may be restricted to
specific tasks. Both methods offer strong performance with some limitations.

Dept. of AI&DS2024 - 2025 Page | 5

Ship Detection from Satellite images using Deep Learning

CHAPTER 4: PROBLEM STATEMENT & OBJECTIVES

Problem Statement:

The growing demand for efficient object detection in satellite imagery presents significant
challenges, particularly in environments with limited labeled data. Traditional object
detection models often rely on large annotated datasets, which are resourceintensive and
timeconsuming to compile. Additionally, these models may struggle to generalize across
different datasets or handle computationally expensive tasks. This project aims to address
these challenges by exploring more efficient detection techniques that reduce the need for
large labeled datasets while maintaining high accuracy and performance.

Objectives:
• Develop an efficient object detection model that minimizes the need for labeled data.
• Explore the use of compact neural network architectures for satellite video object
detection.
• Achieve high detection accuracy while reducing computational complexity.
• Investigate the applicability of active learning strategies to further reduce labeled
sample requirements.
• Assess the performance of the developed model on various datasets with different
characteristics.

Dept. of AI&DS2024 - 2025 Page | 6

Ship Detection from Satellite images using Deep Learning

CHAPTER 5: SYSTEM DESIGN

5.1. System Design:

The proposed system for efficient object detection in satellite imagery will leverage a hybrid
approach combining compact neural networks and active learning techniques. The system will
consist of the following key components:

1. Data Collection and Preprocessing:

Satellite video data will be collected and pre-processed, including normalization and annotation
of relevant objects (e.g., vehicles). This step ensures that the data is ready for training the model
while reducing the volume of labelled data needed.

2. Object Detection Model:

The core of the system will be a compact 3×3 Convolutional Neural Network (CNN) designed
for efficient object detection. The model will use Non-Maximum Suppression (NMS) for post-
processing, which improves detection accuracy by eliminating redundant predictions.

3. Active Learning Strategy:

To further reduce the need for labeled data, an active learning component will be integrated.
This strategy will intelligently select the most informative samples for labeling, minimizing
the total number of labeled samples required for training while maintaining high model
performance.

4. Training and Evaluation:

The model will be trained on a reduced dataset, continuously improved using active learning.
Performance will be evaluated through metrics like detection accuracy, F1 score, and
computational efficiency, ensuring that the model meets the project’s objectives.

5. Deployment and Testing:

Once the model is trained and optimized, it will be tested on various satellite datasets to validate
its generalizability and robustness. The system will be evaluated for real-time application
potential in satellite video analysis.

Dept. of AI&DS2024 - 2025 Page | 7

Ship Detection from Satellite images using Deep Learning

5.2. CNN Architecture

5.3. Hardware and Software Requirements

High-Performance GPU: NVIDIA GPUs such as RTX 3060/3080 or higher for deep learning
model training.
RAM: Minimum 16 GB, preferably 32 GB or more for efficient processing.
Processor: Multi-core processor, e.g., Intel i7/i9 or AMD Ryzen 7/9.
Storage: SSD with at least 1 TB for datasets, pre-trained models, and software.
Display and Peripherals: Monitor, keyboard, and mouse for user interaction.
Operating System: Windows 10/11, Linux (Ubuntu recommended), or macOS.
Programming Language: Python (preferred for deep learning tasks).
Deep Learning Frameworks: TensorFlow, PyTorch, or Keras.
Image Processing Libraries: OpenCV, PIL (Pillow).
Development Tools: Jupyter Notebook, Anaconda, or IDEs like PyCharm.
Dataset Tools: LabelImg or similar for dataset annotation and preprocessing.
Visualization Tools: Matplotlib, Seaborn for analysis and visualization of results.

Dept. of AI&DS2024 - 2025 Page | 8

Ship Detection from Satellite images using Deep Learning

CHAPTER 6: CONCLUSION

This project successfully addresses the challenge of efficient object detection in satellite
imagery by utilizing a compact 3×3 CNN and integrating active learning strategies. By
reducing the need for extensive labeled datasets, the system achieves high detection accuracy
while maintaining low computational complexity, making it ideal for real-world applications
where labeled data is limited. The use of Non-Maximum Suppression (NMS) further enhances
detection precision, ensuring accurate results in complex satellite video environments.

Furthermore, the active learning approach optimizes the model's training process by selecting
the most informative samples, further minimizing the amount of labeled data required. This
method not only improves the model's performance but also significantly reduces resource
consumption. The outcomes of this project provide a scalable, efficient solution for satellite
object detection tasks, with potential applications in areas such as surveillance, environmental
monitoring, and disaster response.

Dept. of AI&DS2024 - 2025 Page | 9

Ship Detection from Satellite images using Deep Learning

CHAPTER 7: REFERENCES

[1] Pflugfelder, Roman, Axel Weissenfeld, and Julian Wagner. "Deep vehicle
detection in satellite video." arXiv preprint arXiv:2204.06828 (2022).

[2] K. Uehara, H. Nosato, M. Murakawa and H. Sakanashi, "Object Detection in

Satellite Images Based on Active Learning Utilizing Visual Explanation," 2019
11th International Symposium on Image and Signal Processing and Analysis
(ISPA), Dubrovnik, Croatia, 2019, pp. 27-31, doi: 10.1109/ISPA.2019.8868462

Dept. of AI&DS2024 - 2025 Page | 10