世 光 · 世 应 实 验 室
UN TECH OVER 2025 · CHALLENGE 2
The build this lab came out of
In June 2025 we entered a hackathon challenge about stacking hazard data. We won it. The tool we built is not the tool we are building now — and the honest account of what it did and did not establish is more useful to you than the trophy.
The challenge
What was being asked
Challenge 2 of the UN Tech Over Hackathon was posed as "Solving the Geo-Puzzle": the problem of making heterogeneous geospatial hazard layers legible together. Different sensors, different resolutions, different update cadences, different reliability — and a decision-maker who needs to know what is true about one place at one moment.
Our entry, Advanced Multi-Hazard Data Overlay with Children Vulnerability Assessment, was submitted on 17 June 2025 by Zhijun He (lead developer and data scientist) and Tiago Maluta, Innovation Manager at the Lemann Foundation and SEESALT consultant, who led the GIS and vulnerability-assessment side. It was named a global champion.
What it does
Inputs, method, output
Inputs
Satellite products from MODIS, VIIRS and CPC, fused by a multi-sensor engine that weights each source by an automated quality score rather than treating them as equally reliable. Sources disagree; the engine is built on the assumption that the disagreement is information rather than noise to be averaged.
Method
A composite vulnerability index with published weights, run alongside spatial autocorrelation analysis, temporal trend detection and anomaly identification.
The weights are published because they are arguable. Every one of these five numbers is a judgement, and someone who works on child heat vulnerability may think thermal stress at 30% is too high or too low. We would rather have that argument than hide the parameter inside a model.
Output
Interactive multi-layer maps with paired temperature and risk heat maps, analysis charts, generated dashboards for decision-makers, and a Telegram alert bot for low-bandwidth mobile delivery. Stack: Python, NumPy, Rasterio, Folium/Leaflet, Matplotlib, Telegram Bot API, with GeoTIFF and JSON throughout.
See it yourself.
Code — github.com/maluta/un-open-source-week-2025-challenge-2
Live dashboard — Team
SEESALT Summit Dashboard
Why it matters for what we do now
The line is not straight, and we are not going to draw it straight
The original purpose was child-specific climate risk: a monitoring and early-warning tool for humanitarian users. That is not what Adaptation Lab is building. What carries over is the machinery, not the purpose.
Measuring a state's prevention capacity from outside requires solving exactly the problems that project had to solve: fusing independent sources of uneven quality, scoring their reliability explicitly instead of assuming it, and reading several layers together for one place at one time. Evidence of prevention does not appear in what a state reports about itself — it appears in land-use change, infrastructure condition, hazard history and exposure. The overlay is the measurement floor that makes the response-versus-prevention distinction reachable at all.
The one design decision that transfers most directly is the smallest one: giving data reliability a fifth of the index weight. An instrument for fragile states has to treat "we do not know" as a finding, not as a gap to be smoothed over. That principle is now the reason our diagnostic reports intervals rather than scores — see the method.
What the result did not establish
Stated rather than left to be inferred
It was an open competition, judged on a prototype, in a weekend format. Four things it is not:
- Not a partnership. It is not an implementing agreement or a procurement. We hold no partnership or implementing agreement with any UN agency or multilateral institution.
- Not a UNICEF relationship. The submission named UNICEF field workers among intended users. That described who the tool was designed for. It described no relationship with UNICEF, and there is none.
- Not adoption. The code has been public since June 2025 with no measured downstream use — no known external deployments, no substantive issue traffic. Views on a submission page are not use, and we are not going to report them as if they were.
- Not validation of the current work. It is evidence that the engineering is real. It says nothing about whether the diagnostic we are now building is wanted, or possible.
Two claims in the original submission we would not make today: "production-ready, 95% complete" and "the first children-focused climate vulnerability assessment system". The first is hackathon register. The second is very likely false — UNICEF's own Children's Climate Risk Index exists, and our own submission referenced the CCRI country framework. Left here rather than quietly dropped.
If you want to pick it up
The repository is the job description
The remote-sensing role we are recruiting for is, concretely, the work of taking this from a competition prototype to something reusable. The hardest part is not modelling. It is testing whether prevention effort leaves an observable proxy at all — maintenance signatures, monitoring-series continuity, enforcement traces in high-risk zones.
The repository is public, the weights are stated, and the failure modes are written down. You can decide whether the problem interests you without talking to us first. If it does: zhijun.he@yale.edu. You do not need to be at Yale, or a student anywhere. We cannot offer salary, and it is better to say that now.
Questions, or corrections to this account
Zhijun He — zhijun.he@yale.edu
Overview · Premises · Method · AAG 2027
播 种 光 明 ， 收 获 永 恒