Build a Robust Image Classifier for a Climate-Tech Satellite Startup
Overview
What this challenge is about.
You receive a labeled dataset of about 25,000 Sentinel-2 patches (positive = illegal construction visible, negative = not). The dataset is split by region AND by season so you can measure distribution shift honestly. Retrain a ResNet-50 (or a small ConvNeXt if you prefer) using modern augmentation (RandAugment, mixup, cutmix), train with focal loss to handle the 10:1 class imbalance, and add test-time augmentation. Compare to the incumbent on overall accuracy, recall on positives, and per-shift performance (cloudy vs. clear, summer vs. winter). Deliver the robustness report to the ML lead.
The Brief
What you'll do, and what you'll demonstrate.
Retrain the coastal-construction classifier so per-shift recall on positives crosses 80% without crashing overall precision.
Earning criteria — what you'll demonstrate
- Apply modern data augmentation and regularization to a real classification task
- Handle class imbalance with appropriate loss functions and sampling
- Measure distribution-shift robustness explicitly, not just headline accuracy
- Deliver a model + report a working ML team can actually adopt
Program Fit
Where this fits in your program.
Sharpens the same skills your degree expects you to demonstrate.
Skills
Skills you'll demonstrate.
Each one shows up on your verified credential.
Careers
Roles this prepares you for.
Real titles. Real skill bridges. Pick the one closest to your trajectory.
Machine Learning Engineer
Improving a production classifier on real distribution shifts and shipping a drop-in inference script is the day-one work of a junior MLE on any applied team.
This challenge sharpens
- pytorch
- model-evaluation
- data-augmentation
Computer Vision Engineer
Satellite-imagery classification with seasonal robustness is a clean CV-engineer skill bridge; the augmentation recipes transfer directly.
This challenge sharpens
- data-augmentation
- deep-learning
- robustness