[Good First Issue] Build regional bias audit framework for model fairness

[Oshgig]

Overview

ClimateVision's U-Net model may perform differently across geographic regions due to varying forest types, cloud patterns, and training data bias. We need a systematic fairness audit to ensure NGOs in the Congo Basin receive equally reliable predictions as those in the Amazon.

Scope

• Create src/climatevision/governance/bias_audit.py with:
  • run_bias_audit(model_path, regions, metric) → returns fairness report
  • Metrics: demographic parity, equalized odds, predictive parity
  • Support for per-region IoU and F1 comparison
• Implement scripts/audit_model.py CLI tool for running full bias audits
• Create notebooks/07_bias_audit.ipynb demonstrating regional disparity analysis
• Add automated bias gate: CI fails if fairness score < 0.85 across regions
• Generate JSON + PDF bias reports for stakeholder transparency

Acceptance Criteria

• Bias audit runs across Amazon, Congo, and Southeast Asia test sets
• Fairness metrics are computed and stored per model version
• CI blocks releases with unacceptable regional disparity
• Report format is documented and reproducible

Resources

• src/climatevision/models/unet.py — model architecture
• src/climatevision/utils/metrics.py — existing metrics
• Fairlearn library (optional): https://fairlearn.org/

Difficulty: Intermediate
Owner: Linda Oraegbunam (@obielin)
Labels: good first issue, governance, fairness, backend

[Oshgig]

Closing for now — will reopen when Linda's pipeline is finalized.

[Goldokpa]

Reopening now that Linda's pipeline is set up. @obielin — this is ready for you to work on.