Experimental prototype · New
Connected Changes
Discover when unusual environmental conditions across cities may be part of a bigger regional story.
When changes across cities tell a bigger story. Grouping is statistical — not a confirmed footprint, cause, or exposure map. Unusual ≠ unsafe.
Potential regional events
4
Cities involved
10
Population in flagged cities
53 Million
No double counting
Most common category
Air
Interactive regional map
Translucent outlines are grouping aids, not an affected area. Marker size tracks severity, colour tracks category.
Regional events
24 flagged cities stay ungrouped — nearby is not enough.
Hangzhou and 2 nearby air cities (China)
3 cities · 16 Million living in flagged cities · earliest onset 2026-10-05 · 4 day span in this snapshot
Regional 30-day trend
One colour per city. Daily values for the last 30 days, not the 7-day sparkline window. Replay walks the dated observations — not a reconstructed cluster history.
Cursor 2026-10-08 · 3 of 3 cities have an unusual reading on this date in the stored series
Evidence
Observed facts, statistical associations, and unverified explanations are labelled separately.
statistical
Graph grouping
3 flagged cities share category Air, direction, and at least one of: onset timing, series similarity, or (for fire) coincident FIRMS detections. 3 qualifying edges. Geography alone was not enough.
observed
Earliest dated onset in this snapshot
Hangzhou 2026-10-06 · Kunshan 2026-10-05 · Shaoxing 2026-10-06
statistical
Shared modeled fields
3 of 3 members use modeled Open-Meteo / CAMS fields. Nearby cities can inherit the same coarse grid cell. That is an association, not independent confirmation.
observed
NASA FIRMS FRP (50 km buffer)
Hangzhou 149 MW · Kunshan 26 MW · Shaoxing 127 MW. FIRMS is satellite-observed, but cities within ~100 km can share the same 50 km hotspot buffer.
observed
10 m wind speed (direction not stored)
Catalog wind speeds: Hangzhou 12.8 km/h · Kunshan 10.2 km/h · Shaoxing 12.9 km/h. Wind direction is not in the daily catalog, so transport is not inferred.
observed
Local news already stored on city pages
Hangzhou: Hangzhou Air Quality Index (AQI) and China Air Pollution (IQAir) · Hangzhou: Understudied Health Risks of High-Impact Pollution Weather: The Case for Revising Short-Term PM (Science Partner Journals) · Hangzhou: Acute effects of air pollutants on influenza-like illness in Hangzhou, China (Nature) · Shaoxing: Kizilsu Air Quality Index (AQI) and China Air Pollution (IQAir) · Shaoxing: ) and Ground-Level Ozone in the Summer of Southern China (AGU Publications) · Shaoxing: Chengdu Air Quality Index (AQI) and China Air Pollution | IQAir Hong Kong (IQAir)
Hypotheses · unverified
- Hypothesis (unverified): nearby FIRMS hotspots could be contributing particles, or they may be unrelated fires inside the 50 km buffer.
- Hangzhou: Hangzhou Air Quality Index (AQI) and China Air Pollution · IQAir
- Hangzhou: Understudied Health Risks of High-Impact Pollution Weather: The Case for Revising Short-Term PM · Science Partner Journals
- Hangzhou: Acute effects of air pollutants on influenza-like illness in Hangzhou, China · Nature
- Shaoxing: Kizilsu Air Quality Index (AQI) and China Air Pollution · IQAir
- Shaoxing: ) and Ground-Level Ozone in the Summer of Southern China · AGU Publications
- Shaoxing: Chengdu Air Quality Index (AQI) and China Air Pollution | IQAir Hong Kong · IQAir
Clustering does not draw a confirmed affected geographical footprint. No pollution-transport claim is made: wind direction is unavailable in this catalog. Shared CAMS / Open-Meteo cells are not treated as independent sensors. Satellite tiles are not fetched for this prototype; only already-stored FIRMS / GIBS URLs are listed. Multi-day cluster history is unavailable from a single snapshot.
Methodology, assumptions, limitations
Linked from the header as Connected Changes · New. The URL is still /experimental/connected-changes. It does not change production detection.
Algorithm: category-specific graph clustering. An edge requires the same pillar and direction, Haversine distance inside a category radius, and either onset proximity, Pearson similarity of the stored primary series, or coincident FIRMS detections for fire. DBSCAN-style density is expressed as a fixed radius plus connected components — more explainable than a black-box clusterer.
Radii: Air 280 km, Heat 550 km, Rain 350 km, Fire 220 km. Trend similarity at long range is ignored; CAMS fields are often synoptically similar across a continent.
Cluster formation, expansion, and resolution are not reconstructed. This prototype has a single daily catalog snapshot. City onset/peak dates come from each city's current dated series, not from past cluster catalogs.
Wind direction is not stored — only 10 m wind speed. Satellite imagery is not fetched here; only already-stored FIRMS FRP and GIBS URLs are listed. Shared coarse model cells are never treated as independent confirmation.