TrustworthySatGeneration

Inspiration

Generating realistic satellite images from maps has immense potential in geospatial analysis and environmental monitoring. However, the misuse of synthetic data can lead to unethical applications, making it crucial to have robust forensics to identify generated content. This project bridges the gap between innovation and ethical considerations in the field of generative AI.

What it does

Our interactive generative AI web application performs conditional generation of satellite images from maps, allowing users to create realistic satellite imagery that accurately represents various features such as buildings, roads, green areas, and water bodies. Also, it incorporates a deep learning-based forensic framework capable of distinguishing between real and generated satellite images to ensure the ethical and secure use of synthetic data.

How we built it

We used Python, PyTorch, and OpenCV to train the GAN and develop the forensic detection framework, utilizing an NVIDIA A100 GPU equipped with 80GB memory for efficient training and high-performance computing. The web application is built using HTML, CSS, and Streamlit to provide a user-friendly browser interface. We integrated SendGrid for an alert system that notifies users via email if a generated image is detected.

Challenges we ran into

We used Python, PyTorch, and OpenCV to train the GAN and develop the forensic detection framework, utilizing an NVIDIA A100-PCI GPU equipped with 80GB memory for efficient training and high-performance computing. The web application is built using HTML, CSS, and Streamlit to provide a user-friendly browser interface. We integrated SendGrid for an alert system that notifies users via email if a generated image is detected.

Accomplishments that we're proud of

We are proud to have successfully generated high-quality satellite images that can interpret complex objects such as buildings, roads, green areas, and water bodies. Completing the training process for the GAN, which took around 10 hours, was a significant milestone. Furthermore, we developed a high-performing deep learning model capable of accurately detecting generated images, contributing to research in forensic image detection.

What we learned

We learned a great deal, starting from preprocessing image data to managing GPU computation, and building robust deep learning models. Developing the web application using Streamlit was a valuable experience. Most importantly, we gained a deeper understanding of how complex yet powerful generative AI can be when applied to real-world scenarios.

What's next for TrustworthySatGeneration

We aim to collaborate with researchers to expand the dataset and develop a publicly available application that is both user-friendly and secure from adversarial attacks. Our goal is to address the shortage of remote sensing data and contribute to the public, research, and governmental sectors, as well as autonomous systems, by providing insights into geographical patterns and historical landmarks.

Built With

• anaconda
• convolutionalneuralnetworks
• css
• cuda
• cv2
• deeplearning
• google-maps
• html
• jupyterlab
• linux
• python
• pytorch
• scikit-learn
• streamlit