Project - Oil Spill Detection System with Synthetic Data Generation Pipeline and Dynamic Perceiver Architecture model 🌊🛢️
- Project Group - 20
- Student - Sweta Pattnaik
- Student ID - A0296537M
- Department - AIS FT 06 batch, ISS, NUS, Singapore
This project aims to develop an AI-based Oil Spill Detection System using Synthetic Aperture Radar (SAR) images. The system enhances oil spill detection accuracy by leveraging synthetic data generation and the Dynamic Perceiver architecture for efficient image analysis.
✅ Detection of oil spills from SAR images
✅ Synthetic data generation using cDDPM (Conditional Denoising Diffusion Probabilistic Model)
✅ High-accuracy model (>90%) using Dynamic Perceiver
✅ Web-based interface
✅ Optimized for Edge AI deployment
- Original SAR Dataset: CSIRO Sentinel-1 SAR Image Dataset
- Synthetic Data: Generated using cDDPM
- Combined Dataset: Real + Synthetic images
- Data Collection & Preprocessing
- Acquire real SAR images
- Generate synthetic SAR images
- Model Training
- Train mobilenet_v3_small as a baseline
- Implement Dynamic Perceiver for improved accuracy
- Evaluate performance using IoU, Precision, Recall, FID
- Web-Based System
- Deploy an interface for uploading SAR images
- Display oil spill detection results
- Edge AI Optimization
- Improve inference speed and reduce power consumption
The detailed project proposal is included in the repository:
📄 ISY5004_ITSS_GC_project_proposal.pptx
The detailed project report is included in the repository:
📄 ISY5004_ITSS_GC_project_report_team_20_Sweta_Pattnaik.pdf
-
CSIRO Sentinel-1 SAR Image Dataset -
Blondeau-Patissier, D., Schroeder, T., Diakogiannis, F., & Li, Z. (2022).
CSIRO Sentinel-1 SAR image dataset of oil- and non-oil features for machine learning (Deep Learning).
DOI: 10.25919/4v55-dn16 -
SAR Image Synthesis with Diffusion Models -
Research paper evaluating DDPM vs. GANs for SAR image generation.
Arxiv Preprint -
Dynamic Perceiver for Efficient Visual Recognition -
A new adaptable Perceiver-based model for image recognition.
GitHub Code Repository -
YOLOv4 for Oil Spill Detection -
Using You Only Look Once (YOLOv4) trained on Sentinel-1 SAR images.- Copernicus Sentinel-1 Data for Oil Spill Detection
- Sentinels Copernicus Case Study
-
Singapore Oil Spill Incident, June 2024 -
- CHATGPT and other AI tools WERE used in this project.