Aesthetic Images and AI: What Does the Machine See?

AI models can describe what is in an image, but can they judge it the way an expert does? For satellite images, a scientist rarely asks one question. Is it beautiful? Is it scientifically interesting? Could it show something harmful, like a pollution plume? A single "aesthetic score" misses all of this.

In this project, you would explore whether a general pretrained vision-language model (such as CLIP) can learn to see satellite images the way one specific expert does. The expert labels a small number of images and writes a short explanation for each. The idea is that the model learns to give a score and to say why, without needing thousands of examples.

You would:

Look at how machines have been taught to judge image aesthetics over the past twenty years

Build a small set of satellite images labeled by an expert

Fine-tune a model on it and compare it with a generic aesthetic model

If time allows, test whether the same idea works for sound

You should have some experience with machine learning or computer vision and be curious about where AI, science, and creativity meet.