SUEWS Hackathon — Practice Repo

This is a practice repository for the SUEWS Community Hackathon, run against the real challenge dataset, UMEP-dev/uda-city-hackathon, which is public ahead of/at kickoff. This is a rehearsal of the pipeline and the analysis approach, not a hackathon submission — the judged entry is a separate repository created under the UMEP-dev organisation on the day.

In one line: the hottest neighbourhoods in this city are not the highest-risk ones — low-income, densely populated districts rank highest on heat risk despite being cooler than the leafy periphery, because risk here is driven by who can't escape the heat, not by how hot it gets. See the chart below.

Pipeline smoke test

Before the real analysis, a minimal SUEWS run via supy (the Python interface suews-agent wraps) confirmed the tooling works end to end, using supy's own bundled sample dataset — see analysis/smoke_test.py and its output.

The question

Across UDA-city's ten neighbourhoods, is the hottest place also the highest-risk place? We ran SUEWS for the present hot-humid season and a +2.5 °C hotter-future pseudo-warming, derived a dangerous-heat-hours hazard layer, and bridged it to a socio-economic heat-risk indicator using the dataset's reference bridge (risk_bridge.py), following a UNDRR-style decomposition:

risk = (hazard × exposure × vulnerability) ^ (1/3)

each pillar min–max scaled to [0, 1] across the ten neighbourhoods. Hazard is dangerous-heat hours (hourly-mean 2 m air temperature, T2, above 35 °C, illustrative threshold, after a 14-day spin-up discard). Exposure is daytime population density. Vulnerability combines five proxies: elderly/under-5 share, lack of AC access, outdoor workers, and a deprivation index.

Method

• Model: SUEWS v2026.6.5 (via supy), canonical config uda-city.yml — NARP net radiation + classic OHM storage heat, 10 grids run in parallel.
• Forcing: ERA5-derived hourly point series for a coastal, hot-humid setting (2024-03-02 to 2024-06-01 local; March is spin-up, the April–May window is analysed).
• Scenarios: present (direct ERA5) and future (+2.5 K uniform pseudo-warming, RH held constant, longwave-down scaled for grey-body consistency).
• Anthropogenic heat is off in both scenarios; population density feeds exposure and vulnerability, not the model's heat input.

Results

Dangerous-heat hours, present vs. +2.5 °C future (full tables: present, future):

Two findings stand out:

1. The hottest neighbourhoods are not the highest-risk ones. The refuge neighbourhoods (Jade Gardens, Taman Melati) post the most dangerous-heat hours in both scenarios — low building roughness means weaker turbulent mixing and warmer near-surface air under NARP — but with zero exposed daytime population they rank last on risk. The dense informal hotspot neighbourhoods (Kampong Lama, Dhobi Lines, Fuzhou Lanes) have markedly fewer hazard hours yet rank highest risk, driven by exposure (300 people/ha) and vulnerability (low AC access, high outdoor work, high deprivation).
2. Warming amplifies the hazard sharply but barely reshuffles the risk ranking. Dangerous-heat hours rose 4–80× across neighbourhoods under +2.5 °C (e.g. Kampong Lama 42→249 hours), yet the risk-rank order is essentially unchanged — because exposure and vulnerability, not hazard, are what separate the top and bottom of the table here. That stability is itself informative: it suggests near-term adaptation priority (who needs protecting first) is robust to this warming scenario, even though the absolute hazard everyone faces is far higher.

What's actually driving this

Three separate mechanisms stack to produce the ranking above, and they don't all point the same way:

• Hazard ← morphology, not vegetation. Refuge neighbourhoods have the lowest building footprint and roughness in the dataset (plan-area fraction λp 0.047–0.073, frontal area index λf 0.02–0.07) — low roughness means weak turbulent mixing, so near-surface air stays warmer under NARP, despite refuge having the most vegetation and water cover of any neighbourhood type (≈26% combined vs. ≈7% for hotspots). Surface greenery does not translate into lower hazard at this grid-average scale.
• Exposure ← population density, not building density. Across the ten neighbourhoods, population density correlates strongly with the final risk score (r ≈ 0.73 — though this is partly definitional, since population is the exposure pillar). Building footprint density does not (r ≈ 0.12) — denser-built neighbourhoods aren't meaningfully riskier here. The two kinds of "density" are easy to conflate and behave completely differently in this model.
• Vulnerability ← adaptive capacity, not just exposure. Holding population density roughly constant (hotspot ≈300/ha vs. core ≈250/ha) isolates the effect: hotspot neighbourhoods have ~10x worse scores on AC access (0.06–0.10 vs. 0.70–0.78), outdoor work (60–65% vs. 18–22%), and deprivation (0.80–0.85 vs. 0.25–0.30). That gap — not hazard, not even exposure — is what actually separates hotspot (rank 1–4) from core (rank 5–6). Deprivation alone correlates at r ≈ 0.87 with risk, but treat that as the formula working as designed (deprivation is a direct input to vulnerability), not an independent discovery.

A scaling artifact worth naming explicitly. The bridge min–max scales each pillar across these ten neighbourhoods only, then takes a geometric mean. Whichever neighbourhood is the dataset's minimum on any single pillar gets scaled to exactly 0, which zeroes the whole risk index regardless of the other two pillars. That is mechanically why refuge (population minimum) and Zheng He Towers (hazard minimum) both land at risk = 0 — it reflects the scaling method, not a claim that those residents face zero real risk.

Two checks the model itself can't run, done with outside information:

• Day vs. night. Every dangerous-heat hour, in both scenarios, falls inside 06:00–18:00 — zero occur at night, even after the uniform +2.5 °C warming. Checking the hourly forcing directly (future scenario): the daytime peak reaches ~35.5 °C around 13:00, but the overnight trough sits in a stable 28.4–29.7 °C band from ~19:00 to ~06:00 — consistently 6–7 °C below the threshold, every night in the record.

  That ~6–7 °C peak-to-trough swing is small for a tropical climate, and it's small for a specific physical reason: humidity. At 81% mean RH, the air carries a lot of water vapour, which is a strong absorber and re-emitter of longwave radiation. During the day, that same humidity (plus the cloud cover that tends to accompany it) reflects and scatters some incoming solar radiation, capping how hot the surface gets. At night, the effect runs the other way and works against cooling: water vapour absorbs the longwave radiation the ground is trying to radiate away to space and re-emits a fraction of it back down, which slows the rate of nighttime radiative cooling considerably compared to a dry climate. A desert location can lose 15–20 °C overnight because dry air barely impedes that radiative loss; this humid coastal setting loses only ~6–7 °C for the same reason in reverse — both the daytime ceiling and the nighttime floor get compressed toward the middle by the same humidity.

  It is not a latitude or night-length effect — at 6.93°N the night is a stable ~12 hours year-round (no seasonal compression the way higher latitudes get in summer), but that duration isn't what's doing the work here; the limiting factor is the rate of cooling per hour, not how many hours are available to cool. The practical upshot: the city retains full overnight cooling relief in both scenarios despite the narrow margin, and it validates using daytime population for exposure, since that's who's present when the hazard actually occurs.

• Ventilation. The forcing has no wind direction (wdir is entirely missing), and this SUEWS configuration is direction-blind by design (λf is an isotropic average) — it cannot represent street-canyon channelling or how building layout interacts with the real-world SW-monsoon-driven prevailing wind a Colombo-like coastal setting would have. Applying general urban-canopy ventilation criteria to our own λf values as an independent check: Mlima Moto (λf=0.90) and Fuzhou Lanes (λf=0.59) fall into a "skimming flow" regime associated with poor street-level ventilation — both already rank among the highest-risk neighbourhoods, so this is corroborating evidence from outside the model, not a SUEWS result.

Where this bridge holds, and where it breaks

• SUEWS gives an environmental hazard, not a health outcome. T2 over 35 °C is a proxy for dangerous conditions, not a prediction of heat-related illness or death.
• 35 °C dry-bulb is a debatable threshold for a humid city (mean RH ≈ 81% here). A humid-heat index or wet-bulb/apparent-temperature metric would more honestly capture physiological danger; we used the dataset's illustrative default rather than re-deriving one, which is itself a limitation worth flagging.
• The socio-economic layer is synthetic — plausible magnitudes for a low-income tropical city, not survey data for a real place. Treat the ranking as the meaningful signal, not the absolute risk-index values.
• Min–max scaling is relative to these ten neighbourhoods only and does not transfer to other cities or datasets.
• Neighbourhood-level aggregation hides intra-neighbourhood variation — a district-mean vulnerability score can mask the most exposed individuals within it.
• The future scenario is a uniform-delta pseudo-warming (+2.5 K, RH-preserving), not a downscaled climate projection — a controlled "what if" stress test, not a forecast.
• The geometric-mean combination is a deliberate choice: a near-zero pillar (e.g. no exposed population) pulls risk to zero even under extreme hazard. An arithmetic mean would let one high pillar dominate instead — equally defensible, but not what we used here.
• The "hotspot" label doesn't match its implied building form. Classifying each neighbourhood's Local Climate Zone (Stewart & Oke, 2012) from λp and mean building height, the four hotspot neighbourhoods — described as "dense informal settlements" — have footprints of only 0.14–0.35, well below the 0.6–0.9 that LCZ 7 (lightweight low-rise, the standard informal-settlement signature) would require. Their nearest LCZ is actually 9 (sparsely built) or the LCZ 6/8 boundary. The heat/risk signal we found for these neighbourhoods comes from who lives there (exposure, vulnerability), not from informal-settlement-style building density — worth being precise about so the result isn't misread as a building-morphology finding.
• No wind direction, and a direction-blind model (see above) — real ventilation differences between neighbourhoods of identical bulk roughness are invisible here.

How this could inform risk reduction

Mapped to the three pillars that actually drive the ranking, rather than generic heat advice:

• Vulnerability is the highest-leverage target. The ~10x gap between hotspot and core on AC access, outdoor-work exposure, and deprivation — not hazard — is what separates their risk ranks. Cooling-centre access, shifting outdoor work away from the 11:00–15:00 peak (when the dataset's own diurnal data shows danger concentrated), and deprivation-focused support target the mechanism we actually found, not just the hottest map pixel.
• Pedestrian-level shade, even though the grid-average hazard didn't show a vegetation effect. SUEWS's neighbourhood-average T2 found refuge's extra vegetation didn't lower its hazard — but that's an average over the whole grid cell. Street-level shade and high-albedo surfaces in the hotspot neighbourhoods would cool conditions where people actually stand, at a resolution this model doesn't resolve.
• Street-layout / ventilation corridor design in Mlima Moto and Fuzhou Lanes specifically, given their poor-ventilation flagging above — informed by real wind climatology (SW monsoon, May–September) that this dataset doesn't include.
• Not population reduction. Exposure tracks where people already live; the intervention there is siting cooling infrastructure to match existing density, not treating density itself as the problem.

Citing SUEWS

• Järvi, L., Grimmond, C.S.B. & Christen, A. (2011). The Surface Urban Energy and Water Balance Scheme (SUEWS): Evaluation in Los Angeles and Vancouver. Journal of Hydrology, 411(3–4), 219–237. https://doi.org/10.1016/j.jhydrol.2011.10.001
• Ward, H.C., Kotthaus, S., Järvi, L. & Grimmond, C.S.B. (2016). Surface Urban Energy and Water Balance Scheme (SUEWS): Development and evaluation at two UK sites. Urban Climate, 18, 1–32. https://doi.org/10.1016/j.uclim.2016.05.001