Mars Landmark Classifier
Compressing Mars imagery for rover communication
- 98.4%dimensionality reduction, 224×224 images to 500-length vectors
- 84.26%test accuracy classifying from the encodings alone
The motivation was communication: a rover sending full images back to a communication hub is expensive, so I wanted to see how much an image could be compressed while staying useful for analysis on the other end.
I built a custom CNN autoencoder whose encoder compresses 224×224 images into compact 500-length latent vectors, a 98.4% reduction in dimensionality. Landmark classification then runs on those encodings rather than on raw pixels, which tests whether the compressed representation keeps the information that matters.
To balance the dataset, I generated 2,000 synthetic images for underrepresented classes using augmentation.
The model reached 98.98% training accuracy and 84.26% test accuracy. The gap between the two points to overfitting, which is the clearest place to improve next.