[Good First Issue] Add SHAP explainability to segmentation predictions

[Oshgig]

Overview

ClimateVision currently produces segmentation masks with zero explainability. NGOs and government agencies cannot trust black-box predictions for high-stakes environmental decisions. We need pixel-level attribution showing WHICH bands and spatial regions drove each classification.

Scope

• Integrate SHAP (DeepExplainer) into the U-Net forward pass in src/climatevision/inference/pipeline.py
• Create src/climatevision/governance/explainability.py with:
  • explain_prediction(model, image, analysis_type) → returns SHAP values + heatmap
  • Support for all 3 analysis types (deforestation, ice_melting, flooding)
  • Band-level attribution (e.g., "NIR contributed 47% to this forest pixel classification")
• Add GET /api/explain/{run_id} endpoint returning SHAP values as GeoJSON/PNG
• Create notebooks/06_explainability.ipynb with visual examples
• Add pytest tests verifying explainability output shapes match prediction shapes

Acceptance Criteria

• Every prediction can return a SHAP heatmap overlay
• Band attribution is returned in API response
• Notebook demonstrates explainability across all 3 analysis types
• Tests pass in CI

Resources

• src/climatevision/models/unet.py — U-Net architecture
• src/climatevision/inference/pipeline.py — inference flow
• SHAP library docs: https://shap.readthedocs.io/

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