geointel/docs/SPECIFICATION_FREEZE_M0.md

GeoIntel Kempen — Specification Freeze M0

Status: frozen for first Codex build Purpose: remove architecture ambiguity before implementation

1. Product decision

GeoIntel Kempen is a GeoAI Workbench, not a generic GIS viewer and not a policy-report dashboard. The primary value is the conversion of raster, vector and AI outputs into geospatially correct layers, metrics, QA findings and exports.

2. Region decision

The default operating region is the Belgian Kempen, with Flemish open geodata as the professional reference layer. The initial demo area should be Geel, Mol or Turnhout because these are recognisable Kempen cases with buildings, vegetation, water and infrastructure.

3. First-user decision

The first target user is a technical geospatial analyst or GeoAI engineer candidate. UI copy may be understandable to non-specialists, but implementation must prioritise correct geodata handling, reproducible pipelines and inspectable outputs.

4. V1 priority order

1. Project workspace and area selection.
2. Dataset intake and metadata extraction.
3. Raster and vector workbench foundation.
4. GRB/OSM reference-layer ingestion or caching.
5. Object detection pipeline.
6. Segmentation pipeline.
7. QA/QC against GRB or reference vector layers.
8. Export of outputs as GeoJSON, masks and simple report artifacts.

5. Explicit non-goals for V1

• No multi-tenant user management.
• No payment, sharing or organisation management.
• No QGIS plugin yet.
• No full LiDAR workbench yet.
• No full model training studio yet.
• No heavy MLOps platform yet.
• No beautiful but empty dashboard replacing actual processing.
• No fake “analysis complete” states without persisted outputs.

6. Data source decisions

• GRB is the first professional Flemish reference source.
• OSM is the fallback and supplementary open context source.
• Sentinel-2 is the first remote-sensing source for NDVI/NDWI/NDBI, but only after raster/vector foundation is stable.
• DHMV/DEM/DSM are later height sources.
• LAS/LAZ point clouds are later, not V1.

7. AI decisions

• V1 may use pre-trained YOLO/Ultralytics models and SAM-style segmentation where practical.
• The system architecture must isolate model inference behind services so models can be replaced later.
• AI results must be georeferenced back into map coordinates.
• AI results must never be treated as ground truth without QA/QC.

8. QA/QC decisions

The signature V1 portfolio feature is comparing AI-generated detections or segmentations against GRB or another reference vector layer. The QA engine must calculate at least:

• matched features
• unmatched predictions
• unmatched reference objects
• IoU distribution
• precision
• recall
• F1

9. Export decisions

V1 exports must be useful to GIS workflows:

• GeoJSON for vector outputs
• zipped result package with metadata JSON
• mask raster path or exported raster artifact where segmentation is used
• simple HTML or Markdown report summary; advanced PDF can come later

10. Implementation rule

If a choice is not explicitly frozen here, Codex must prefer the simplest implementation that preserves geospatial correctness, testability and later extensibility.