B-MCRI is an independent, open fire-risk map for Karachi. It takes the things that make a fire worse (packed housing, blocked roads, a slow response, a history of past incidents) and turns them into a single score for every block.
Tap any unit to see its structural index R̂, the Bayesian-network posterior, and the season it's in. Boundaries follow the 2001 city-district delimitation: 178 union councils across 18 towns plus 6 cantonment boards. The newer 246-committee structure isn't mapped yet.
No single model sees the whole picture, so B-MCRI stacks four. Each one answers a question the others can't, and together they fold into a single score you can actually read.
Start with what you can see on a map: how tightly the buildings are packed, how a fire truck would reach the street, how far the nearest station is, what the land is used for, and whether it's an informal settlement. Those roll up into four dials (Hazard, Exposure, Vulnerability, and protective Capacity) and combine into a score from 0 to 1. Everything after this is just a correction on top.
Risk factors aren't independent. Old wiring, occupancy, and season all lean on each other. The Bayesian network writes those links as a graph and works out P(Fire | evidence), updating its belief as conditions change and staying sensible even when half the inputs are missing.
Fire risk breathes with the calendar. From the run of past detections, an HMM picks out three hidden moods (Calm, Transitional, Dry), each nudging the structural score up or down. It's how a flat map learns to tell January from a dry, load-shedding July.
Urban fires cluster. One blaze pulls in nearby crews, jumps across shared walls, and briefly raises the odds of another next door. The Hawkes process captures that as a conditional intensity λ(t): a background rate plus a decaying jolt after every incident. Have a go below.
Click the plot (or the button) to start a fire. Watch the intensity jump, then cool back down toward the background rate. That's the same self-excitation the model applies across the city.
Five feeds, each open or public-domain. Some are live; some are frozen at a census epoch. All of it is traceable.
Every model has limitations. These are the main ones, set out plainly rather than buried in a footnote.
B-MCRI started from a gap. Karachi keeps losing people to preventable fires in dense, hard-to-reach settlements, and there is almost no spatial risk data anyone can act on.
It combines classic spatial risk scoring with probabilistic methods (Bayesian networks, Hidden Markov Models, and self-exciting Hawkes processes) into an index that updates as new incident data comes in. The goal is a working tool, not only a paper.
Collaboration, data access, or a gap you've spotted in the method: email is the best way to reach me.
“The point of a risk map isn't to predict the next fire. It's to argue, with evidence, about where the next truck, hydrant, or inspection should go.”