UHI-Bench-Data: Dataset for Benchmarking Urban Heat Island Modeling

UHI-Bench-Data is the released dataset for UHI-Bench, a benchmark for urban heat island (UHI) modeling across data sources, cities, climate regimes, and tasks.

The release aligns land-surface-temperature UHI (LST-UHI), near-surface air-temperature UHI (AirT-UHI), hourly meteorological drivers, and static urban morphology features on per-city 1 km grids. It covers 20 gridded cities across nine Köppen climate classes, plus two supplementary station-format AirT cities.

Total size: ~232 GiB, ~3 000 files.

---

1. Dataset Scope

Gridded cities

16 dual-source core cities provide paired LST-UHI and AirT-UHI fields and support the main benchmark tasks.

• Germany / HOSTRADA AirT reference (8): Berlin, Hamburg, Munich, Cologne, Dortmund, Dusseldorf, Frankfurt, Stuttgart
• International / corrected reanalysis-downscaled AirT (8): Cairo, Johannesburg, Lagos, Riyadh, Bucharest, Warsaw, Sao Paulo, Buenos Aires

4 LST-UHI-only transfer cities provide LST-UHI, ERA5, and static features for surface-UHI transfer evaluation:

• Tehran, Khartoum, Casablanca, Istanbul

Station-format supplementary cities

Rome and Temuco provide station-format AirT observations for additional sparse near-surface temperature anomaly detection and forecasting checks. They are not part of the 1 km gridded dual-source core.

Intended benchmark use

UHI-Bench-Data supports:

• cross-source LST-UHI vs AirT-UHI consistency analysis
• extreme-event detection
• LST cloud-gap imputation
• AirT sparse-grid reconstruction
• multi-horizon UHI forecasting
• meteorological and morphology driver attribution
• cross-city and cross-climate transfer

2. Directory Layout

release/
├── hostrada_uhi/{city}/                 # AirT-UHI reference product for German 8 cities
│   ├── grid_centers.csv                 # pixel_id, lon, lat, x_epsg3034, y_epsg3034
│   ├── monthly_uhi/uhi_YYYYMM.parquet   # datetime, x_epsg3034, y_epsg3034, uhi [K]
│   └── monthly_met/met_YYYYMM.parquet   # tas, sfcWind, rsds, hurs on HOSTRADA grid
├── atuhi_ood_1km_hourly/{city}/         # AirT-UHI for international 8 cities
│   └── atuhi_ood_1km_hourly_YYYY.parquet
│       # datetime, pixel_id, uhi
│   └── qa/                              # optional corrected AirT-UHI vs LST-UHI QA summaries
├── lstuhi_1km_hourly/{city}/            # LST-UHI for 20 gridded cities
│   └── lst_uhi_1km_hourly_YYYY.parquet  # pixel_id, datetime, lst_K, lst_uhi_K
├── static_features/{city}/              # Static and time-varying urban morphology features
│   ├── static_features.npz              # pixel_ids, xy, features, feat_names
│   ├── raster_aligned.csv, derived_features.csv, grid.gpkg
│   ├── temporal_static/{year}.npz       # year/season-varying static layers
├── temporal_weather/{city}/             # ERA5 meteorological drivers on 1 km grids
│   └── era5_hourly_YYYY.parquet         # datetime, pixel_id, t2m, u10, v10, tcc, d2m, blh, ssrd
├── station_uhi/                         # Station-format supplementary data
│   ├── rome_uhi/processed/              # anomaly/timestamp arrays, JJA 2019-2020
│   └── uhi_temuco/processed/            # anomaly/timestamp arrays, 2017-2018
└── README.md

3. Spatiotemporal Alignment

All gridded products are aligned to a fixed per-city 1 km pixel space. Within each city, pixel_id is the key used to align LST-UHI, AirT-UHI, ERA5, and static features. ERA5 variables are bilinearly interpolated to the 1 km pixel centers. MSG/SEVIRI LST is downscaled or resampled to the same grid. AirT-UHI products are mapped to the same pixel IDs before cross-source comparison or modeling.

All hourly source matching is performed in UTC. The parquet datetime columns are timezone-naive timestamps, but they should be interpreted as UTC. ERA5 joins, LST-AirT matching, train/evaluation splits, and benchmark model inputs use exact UTC hourly timestamps. This convention avoids ambiguity around daylight-saving transitions.

For heat-risk timing diagnostics only, extreme labels are first constructed on the UTC-aligned series and then timestamps are converted from UTC to each city's local civil time using the appropriate IANA time zone. Hour-of-day and day/night timing plots therefore report local hour, while the underlying data fusion and model evaluation remain UTC-aligned.

4. UHI Targets

Group	AirT-UHI	LST-UHI
German 8	HOSTRADA/DWD 1 km reference product (hostrada_uhi)	MSG/SEVIRI LST-UHI on the same 1 km grid (lstuhi_1km_hourly)
International 8	Corrected ERA5-constrained reanalysis-downscaled product (atuhi_ood_1km_hourly)	MSG/SEVIRI LST-UHI with 1 km downscaling (lstuhi_1km_hourly)
LST-only transfer 4	Not included	MSG/SEVIRI LST-UHI with 1 km downscaling (lstuhi_1km_hourly)
Station supplementary 2	Station-format AirT observations (station_uhi)	Not included

Both UHI targets are reported in kelvin (K) as urban-minus-rural anomalies. LST-UHI is derived from local LST minus the contemporaneous rural-reference mean and retains cloud-contaminated retrievals as missing values. AirT-UHI is local near-surface air temperature minus a rural-reference mean.

Important caveat: international AirT-UHI is not station truth. It is a corrected reanalysis-downscaled product constructed from ERA5 temperature with a HOSTRADA-trained residual correction model. It should be reported separately from the German HOSTRADA AirT-UHI reference product.

Hamburg grid note: the HOSTRADA AirT-UHI grid for Hamburg is 2 216 pixels. Fourteen edge/water pixels that were permanently NaN in HOSTRADA were removed from grid_centers.csv, monthly_uhi/, and monthly_met/, so Hamburg AirT missingness is 0.00%. static_features keeps the full 2 230-pixel superset; loaders that key off the AirT grid automatically use the aligned 2 216.

5. Temporal Coverage and Benchmark Split

Group	Cities	Range	Default benchmark role
16 dual-source core	German 8 + international 8	2015-01 to 2025-12	Main signal consistency, detection, imputation, forecasting, attribution, transfer
Train/stat period	Core gridded cities	2015-01 to 2022-12	Training years and climatology/statistics years
Evaluation period	Core gridded cities	2023-01 to 2025-12	Held-out evaluation years
4 LST-only transfer cities	Tehran, Khartoum, Casablanca, Istanbul	2019-01 to 2025-12	Surface-UHI transfer and OOD evaluation
Station supplementary	Rome, Temuco	city-specific ranges	Station-format AirT detection and forecasting checks

Per-city grid counts and full-calendar missing rates are reported in the paper appendix tables and associated benchmark metadata.

6. Quick Load Examples

import numpy as np
import pandas as pd

# German AirT-UHI reference product
air = pd.read_parquet("hostrada_uhi/berlin/monthly_uhi/uhi_202401.parquet")

# LST-UHI
lst = pd.read_parquet("lstuhi_1km_hourly/cairo/lst_uhi_1km_hourly_2024.parquet")

# Static morphology features
static = np.load("static_features/berlin/static_features.npz", allow_pickle=True)
features = static["features"]
feat_names = [str(x) for x in static["feat_names"]]

# Meteorological drivers
era5 = pd.read_parquet("temporal_weather/cairo/era5_hourly_2024.parquet")

7. Static Features

static_features/{city}/static_features.npz stores per-pixel morphology and environmental features. The canonical cross-city set is the first 10 Tier-1 features, available for all 20 gridded cities. German cities additionally include 6 German Building Attribute (GBA) features.

feat_names ordering is fixed: Tier-1 features first (indices 0-9), then GBA features (indices 10-15, German 8 only). Benchmark loaders that require cross-city parity should slice features[:, :10].

Tier-1 features, all 20 gridded cities

#	Name	Time behavior	Meaning / source
0	BCR	permanent	Building coverage ratio
1	road_density	permanent	Road density
2	poi_density	permanent	POI density
3	nightlight	annual	VIIRS nighttime lights
4	ndvi	seasonal	Sentinel-2 NDVI
5	water_ratio	annual	Water-body ratio
6	distance_to_waterbody	annual	Distance to water
7	mean_height	permanent	Mean building height
8	dem	permanent	Elevation
9	wind_exposure_proxy	annual	Wind exposure proxy

Additional GBA features, German 8 only

#	Name	Meaning
10	FAR	Floor area ratio
11	volume_density	Building volume density
12	svf_proxy	Sky-view-factor proxy
13	roughness_proxy	Surface roughness proxy
14	std_building_height	Standard deviation of building height
15	hw_ratio	Height-to-width proxy

static_features.npz is a snapshot assembled for release. Permanent features are time-invariant. Annual and seasonal layers, such as nightlight, NDVI, water ratio, and wind exposure, are also provided in temporal_static/{year}.npz where available. Use temporal_static for strict timestamp-specific feature construction.

Benchmark convention: UHI-Bench uses the 10 Tier-1 features by default for cross-city fairness. The 6 GBA features are excluded from cross-city default protocols and can be used only in German-only ablations.

8. Sources and License

Main sources include HOSTRADA/DWD, MSG/SEVIRI LST, ERA5 single-level products, Copernicus elevation and wind-related products, OpenStreetMap/Geofabrik, VIIRS nighttime lights, Sentinel-2 NDVI, JRC Global Surface Water, GlobalBuildingAtlas LoD1, ASTI-Network Rome observations, and Temuco station observations from Martinez-Soto et al. 2023.

The public release ships benchmark-ready gridded products, aligned feature tables and tensors, and processed station anomaly arrays. Raw GEE raster exports, raw OSM/Geofabrik vectors, and raw station source files are not included in the public release.

Derivative compiled files are released under CC-BY-4.0 where permitted. Third-party products retain their original licenses and terms of use.