I’m currently a college student and new to Deep Learning and Computer Vision with a strong interest in applying machine learning to astronomical data analysis. I’m working on a project where I aim to analyze the impact of light pollution on night sky observations using deep learning techniques.
Project Overview
The core idea of my project is to compare the visibility of celestial objects in clear skies versus polluted skies. Specifically, I plan to:
Collect Data: Obtain 24-hour time-lapse videos of the night sky from two locations in India:
Clear Sky: Hanle Observatory (Ladakh)
Moderately Polluted Sky: Guru Shikhar Observatory (Mount Abu)
Analyze and Compare: Use deep learning models to extract features from the videos, such as star count, brightness distribution, and object visibility. My goal is to quantify the “light pollution factor” and understand how pollution affects astronomical observations.
What I’m Seeking
Guidance on Model Selection: Suggestions on the best DL architectures for analyzing time-lapse videos.
Data Preprocessing Tips: Advice on effective techniques for processing and normalizing video data for this type of analysis.
If you have any experience in video analysis, astronomical data processing, or related machine learning tasks, your insights would be invaluable to me. Even suggestions on potential pitfalls or resources to explore would be greatly appreciated.
To me this seems more like an image processing task rather than a deep learning topic.
For this to use deep learning, you will need a large dataset of labeled examples, so that you can create a model and that will allow predictions on new data.
How does this definition of deep learning apply to your task?
Thank you for your feedback! I definitely didn’t think about that. The definition certainly helps me in thinking about my task in a new perspective.
I plan to gather a large dataset of time-lapse videos from different locations with varying levels of light pollution. By extracting frames at regular intervals, I can create a dataset of images that capture different conditions of the night sky.
Each frame can be labeled based on the level of light pollution or the number of visible stars, creating a labeled dataset. I could use both existing datasets and my own collected data to build an extensive dataset.
And I could use CNNs and image augmentation to automatically extract features from the images, such as star patterns, brightness, and other indicators of light pollution.
Since the data is sequential, I could add RNNs to analyze how these features change over time, allowing the model to understand the dynamics of light pollution.
If you have any further suggestions or thoughts on what image processing or deep learning technique to use into this project, I’d love to hear them!