geointel/docs/CHANGE_DETECTION_SPEC.md

GeoIntel Kempen — Change Detection Specification v1.0

Change Detection is a major showcase workflow combining raster, vector, AI and QA/QC.

Goal

Compare two datasets or analysis runs for the same area and detect additions, removals and significant changes.

Supported methods

Method A — Vector change detection V1

Compare two vector layers or two analysis outputs.

Examples:

• GRB buildings snapshot A vs snapshot B.
• AI detections from raster A vs AI detections from raster B.
• OSM buildings from run A vs OSM buildings from run B.

Inputs

• layer A
• layer B
• area polygon
• class filter optional
• matching threshold

Algorithm

1. Normalize CRS.
2. Clip both layers to area.
3. Match features using IoU or spatial overlap.
4. Classify:
   • added: feature in B without match in A
   • removed: feature in A without match in B
   • unchanged: matched with stable geometry
   • modified: matched but area or geometry changed above threshold

Metrics

• added count
• removed count
• modified count
• added area m²
• removed area m²
• net area change m²
• percentage change

Method B — Raster index change V2

Compare NDVI/NDWI/NDBI rasters.

Inputs

• index raster A
• index raster B
• threshold

Algorithm

delta = index_B - index_A
classify pixels by threshold
polygonize changed zones

Outputs

• change raster
• changed polygons
• summary statistics

Method C — AI segmentation change V3

Run segmentation on both images and compare class polygons.

Examples:

• vegetation loss
• new buildings
• water change

Output layers

• change_added
• change_removed
• change_modified
• change_heatmap
• change_uncertain

API

POST /analysis/change-detection
GET /analysis/{id}/changes
POST /analysis/{id}/exports/change-geojson

UI requirements

Change Lab must support:

• dataset/layer A selector
• dataset/layer B selector
• method selector
• area selector
• threshold controls
• timeline labels
• map overlays for added/removed/modified
• metrics cards
• export

V1 target

Implement vector change detection. Raster and AI-based change detection are later phases.