Technical depth / education

Mars surface image classification (CNN features + K-Means)

DCU coursework on the NASA Mars Surface Image Dataset (6,691 labelled Curiosity rover images across 25 categories). EfficientNetB0 for feature extraction, PCA for dimensionality reduction, K-Means for unsupervised clustering, with a silhouette score of around 0.12 to 0.15.

Employer / client: Dublin City University
Duration: EEN1083/EEN1085, 2024
Project type: Computer vision / clustering

GitHub mars-surface-classification

Architecture

How the clustering pipeline works

How the rover images flow through EfficientNetB0 feature extraction, PCA reduction, and K-Means clustering, then get judged on silhouette score and cluster make-up.

My clustering pipeline

Process flow

How I work the steps

01
before Load dataset
NASA Curiosity images, 25 labelled categories.
Public dataset
02
control Extract features
EfficientNetB0 as a frozen feature extractor.
Me
03
handoff Reduce + cluster
PCA to 50 components, then K-Means.
Me
04
after Evaluate
Silhouette ~0.12 to 0.15 plus cluster inspection.
Me

How I built it

EfficientNetB0 used as a frozen feature extractor; PCA reduced the feature vectors to 50 components before clustering.
Evaluated cluster quality with silhouette score and visual inspection of cluster compositions, not just accuracy against the labels.
Code and notebook on GitHub so the analysis is reproducible end to end.

Measured results

What I measured

6.7K evidence

NASA Curiosity images clustered

K-Means on PCA-reduced EfficientNetB0 features produced meaningful groupings for visually distinct surface types (drill, wheels, horizon).

Findings

K-Means on PCA-reduced EfficientNetB0 features produced meaningful groupings for visually distinct surface types (drill, wheels, horizon).
Categories with high visual similarity (e.g. ground vs observation tray) overlapped, as expected. Written up rather than glossed over.
Silhouette score of about 0.12 to 0.15: positive separation, not clean.