Computer Vision (Analyst) · Level 201
Computer Vision Analyst 201 — Trusting the Model
Run EuroSAT evaluation notebooks, read where models fail, and write briefs colleagues can act on — no architecture tuning required.
Computer Vision (Analyst) · Level 201
Computer Vision Analyst 201 — Trusting the Model
Run EuroSAT evaluation notebooks, read where models fail, and write briefs colleagues can act on — no architecture tuning required.
Computer Vision Analyst 201 — Trusting the Model
Prerequisite: Computer Vision Analyst 101.
Developers on your team train models; you evaluate whether colleagues can act on the output. You run provided notebooks (or review worked charts) and write Andy’s park brief.
Want to train MobileNet yourself? → Computer Vision Developer 201.
Before you start
Download
land-cover-analyst-201-starter.zip.
If EuroSAT download fails, use charts on Worked examples.
Module 1 — EuroSAT Samples
Open notebooks/module1_eurosat_samples.ipynb. Run All.
Save outputs/charts/module1_eurosat_samples.png.
In report/module1_notes.md (100–150 words):
- Could you label Pasture vs HerbaceousVegetation without a legend?
- Why might a ranger care about the difference?
- What is missing compared to a full Peak District APGB map?
Module 2 — Read the Confusion Matrix
Open notebooks/module2_confusion.ipynb. Run All.
The notebook prints test accuracy and saves outputs/charts/module2_confusion.png.
Write report/module2_evaluation_note.md (150–200 words):
- Test accuracy — quote the number from the notebook.
- One pair of classes the model confuses — why visually?
- One reason that accuracy overstates real Peak District performance.
How to read the matrix: rows = true class, columns = predicted. Diagonal = correct; off-diagonal = mistakes that could mislead a grazing decision.
Module 3 — Augmentation in Plain English
Read the notebook notebooks/module3_augmentation_notes.ipynb (markdown + one
chart cell). It shows validation curves with and without augmentation.
Answer in report/module3_augmentation_note.md (100–150 words):
- Did augmentation improve validation accuracy on this run?
- In plain English, why might flips and brightness shifts help aerial mapping?
- Why is augmentation not a substitute for wrong labels?
Augmentation may not beat the baseline — honest reporting is the skill.
Module 4 — Transfer Learning in Plain English
Read notebooks/module4_transfer_explainer.ipynb. It loads the saved model and
prints test accuracy.
Answer in report/module4_transfer_note.md (100–150 words):
- Why might ImageNet help on moorland tiles even though ImageNet has no heather?
- What is one thing you would check before deploying on next year’s APGB flight?
- Why must Andy report limitations even at 90%+ accuracy?
Module 5 — The Park Brief
Deliverables:
outputs/charts/module5_gallery.png— fromnotebooks/module5_brief.ipynb
(correct vs incorrect tiles), or copy from worked examples and cite them.report/module5_park_brief.md— 400–500 words for rangers and Landscape Recovery partners.
Structure:
- What we mapped (four EuroSAT classes — teaching proxy, not operational).
- How (high level — a convolutional neural network (CNN) with transfer learning; no code dump).
- Accuracy + worst confusion pair.
- Limitations (geography, resolution, season, no ground survey).
- Recommendation (draft map only; verify pasture on foot before grazing).
See the sample park brief for tone — write your own words.
Next steps: Computer Vision Analyst 301 — your own scoped brief.