UHI-Bench: Benchmarking Dual-Source Urban Heat Island Modeling Across Cities in Diverse Climate Regimes

Wanyun Ling; Chenxi Liu; Yi Xie; Aopu Xu; Zhuoqi Zeng; Ziyue Li

cs.LGAug 24, 2026

UHI-Bench: Benchmarking Dual-Source Urban Heat Island Modeling Across Cities in Diverse Climate Regimes

Authors: Wanyun Ling, Chenxi Liu, Yi Xie, Aopu Xu, Zhuoqi Zeng, Ziyue Li

Organizations: Department of Operations and Technology Technical University of Munich Heilbronn, Germany · Hong Kong Institute of Science & Innovation, Chinese Academy of Sciences Hong Kong SAR, China · Faculty of Computer Science RWTH Aachen University Aachen, Germany · Hainan Bielefeld University of Applied Sciences Hainan, China · Department of Operations and Technology, Heilbronn Data Science Center, Munich Data Science Institute Technical University of Munich Munich, Germany

Abstract

Urban heat islands (UHIs) are intensifying under climate change, exacerbating thermal exposure risks. Their two primary observations, land surface temperature UHI (LST-UHI) and near-surface air temperature UHI (AirT-UHI), capture physically distinct aspects of urban heat. However, most studies rely on a single source, and substituting one for the other can substantially bias the magnitude and spatial variability of human heat exposure. Accurate UHI modeling also requires dynamic meteorological drivers and static urban morphology features, but spatiotemporal incompatibilities hinder their alignment. Cloud gaps in LST observations and sparse AirT station networks further limit dual-source UHI modeling, motivating cross-city transfer across diverse climates. To bridge these gaps, we introduce UHI-Bench, the first UHI benchmark for dual-source UHI modeling that integrates dynamic and static environmental context. Following a unified signal, mechanism, and transfer framework, it evaluates over 20 baselines from four model families on five tasks across 20 cities and nine Köppen climate classes. Results show that no model is uniformly best, although foundation models remain consistently competitive and stable. Environmental covariates generally improve performance, but their utility varies across sources and tasks. Cross-city transferability is better explained by overlap in UHI regimes than by climate-zone similarity. With the dataset and standardized pipeline, our work provides practical guidance for urban heat modeling, promotes climate data equity, and supports future advances in climate research.

Figures & tables

Figure 1. UHI-Bench benchmark overview. The benchmark is organized around three research questions, progressing from signal validity to mechanism understanding and finally to cross-city transfer. Overview graphic summarizing the benchmark motivation, research questions, and tasks spanning signal consistency, detection, imputation, forecasting, attribution, and transfer.

Dataset	Spatial coverage	Year span	UHI data available?	UHI source	Temporal resolution	Spatial resolution	Paired meteorological drivers	Paired static environmental data	Public access
UrbClim ( Lauwaet et al., 2024 )	100 cities	2008–2017	✗	Simulated AirT	Hourly	100 m	✗ a	✗	✓
Global UHII ( Yang et al., 2024 )	10,000+ cities	2003–2020	✓	LST-UHI+AirT-UHI	Monthly	1 km	✗	✗	✓
Global HW Exposure ( Yu et al., 2026 )	Global	2003–2020	✗	LST	Daily	1 km	✗	✗	✓
IBM Johannesburg ( IBM Research, 2023 )	1 city	$\sim$ 10 years (not clearly specified)	✗	AirT	Daily	1 km	✗	✗	✗
US SUHI Database ( Chakraborty et al., 2020 )	497 US urbanized areas	2013–2017	✓	LST	Seasonal	1 km / census tract	✗	✗ b	✗ e
SeoulTemp / DeepUHI ( Zou et al., 2025b )	1 city: Seoul	2021–2024	✗	AirT	Hourly	Street-level; 947 stations	✗	✗ c	✓

Table 1. Comparison between UHI-Bench and related urban heat datasets.

Figure 2. Geographical distribution and climate coverage of the 22-city UHI-Bench Data. Two-panel benchmark geography figure. Panel a maps the 22 cities over global climate zones with a Germany inset. Panel b shows climate-indicator distributions for representative cities.

Figure 3. UHI-Bench data construction pipeline, including temperature downscaling, UHI derivation, and spatiotemporal alignment with urban morphology and meteorological variables. Pipeline diagram showing downscaling of satellite land surface temperature and air temperature, UHI anomaly construction, and spatial-temporal alignment with static morphology and meteorological inputs.

Figure 4. Analysis 1a cross-source consistency across 16 cities. Cross-source diagnostic summary with same-hour correlation, lag instability, best-lag Pearson correlation, and gap-filling gain panels.

Figure 5. Analysis 1b extreme UHI event detection by hour of day and city. City-wise diurnal profiles of AirT and LST extreme-hour shares across the benchmark cities.

		F1 (AirT-UHI\|LST-UHI)
Group	Model	Lagos (Aw)	Johannesburg (Cwb)	Cologne (Cfb)	Riyadh (BWh)
Stat/Geo	Percentile	0.654\|0.248	0.760\|0.621	0.875\|0.268	0.786\|0.329
Stat/Geo	SeasonalNaive	0.371\|0.249	0.508\|0.540	0.496\|0.231	0.462\|0.187
ML	RF	0.740\|0.378	0.739\|0.663	0.939 \|0.261	0.870\|0.481
ML	XGBoost	0.726\|0.361	0.713\|0.633	0.938\|0.226	0.876 \|0.460
ML	RF+m	0.752\|0.396	0.753\|0.659	0.923\|0.294	0.872\| 0.506

Table 2. Task 1 detection leaderboard for Lagos, Johannesburg, Cologne, and Riyadh.

		MAE
		Cairo (BWh, 20.0%)				Bucharest (Cfb, 57.2%)				Lagos (Aw, 75.6%)
Group	Method	0–25%	25–50%	50–75%	>75%	0–25%	25–50%	50–75%	>75%	0–25%	25–50%	50–75%	>75%
Stat/Geo	IDW	1.405	1.568	2.082	2.258	0.989	1.379	1.285	1.440	1.500	1.656	2.005	2.521
Stat/Geo	OrdKrig	1.440	1.623	2.172	2.311	1.007	1.414	1.346	1.471	1.767	1.838	2.161	2.599
Stat/Geo	RegKrig+s	1.255	1.216	1.534	1.746	0.969	1.262	1.205	1.350	1.052	1.125	1.244	1.397
DL	SAITS	1.077	1.124	1.071	1.099	1.843	1.844	1.797	1.818	1.223	1.302	1.355	1.337

Table 3. Task 2a LST cloud-gap imputation MAE for Cairo, Bucharest, and Lagos.

		MAE, by station missing-% bin
		Lagos (Aw)				Cologne (Cfb)				Riyadh (BWh)
Group	Method	0–25%	25–50%	50–75%	>75%	0–25%	25–50%	50–75%	>75%	0–25%	25–50%	50–75%	>75%
ML	RF	0.052	0.053	0.054	0.068	0.1104	0.1188	0.1310	0.1602	0.0776	0.0772	0.0792	0.0933
ML	XGBoost	0.053	0.054	0.055	0.068	0.1061	0.1147	0.1287	0.1595	0.0784	0.0781	0.0796	0.0952
ML	RF+s	0.030	0.030	0.031	0.038	0.0951	0.0995	0.1066	0.1200	0.0525	0.0524	0.0527	0.0644
ML	XGBoost+s	0.027	0.027	0.027	0.034	0.0863	0.0924	0.1017	0.1191	0.0484	0.0483	0.0488	0.0590

Table 4. Task 2b AirT station-sparse imputation MAE for Lagos, Cologne, and Riyadh.

		Avg MAE AirT-UHI				Avg MAE LST-UHI
Group	Method	Lagos	Johannesburg	Cologne	Riyadh	Lagos	Johannesburg	Cologne	Riyadh
ML	XGBoost	0.072	0.135	0.144	0.097	0.789	0.836	0.638	0.808
ML	XGBoost+m	0.064	0.086	0.104	0.063	0.741	0.784	0.615	0.772
ML	XGBoost+s	0.073	0.123	0.144	0.097	0.789	0.834	0.638	0.808
ML	XGBoost+s+m	0.064	0.083	0.102	0.061	0.737	0.767	0.610	0.763
DL	LSTM	0.099	0.193	0.189	0.158	1.012	1.118	0.684	1.276

Table 5. Task 2c forecasting benchmark with average MAE across prediction horizons for Lagos, Johannesburg, Cologne, and Riyadh.

Figure 6. Task 2d Mechanism Stability diagnostics. Two-panel Task 2d mechanism-stability summary with +m driver responses and static-feature responses for AirT-UHI and LST-UHI.

		MAE (RMSE)
Group	Method / source	+1h	+6h	+24h
ML	XGBoost / Cfb-4 (+m)	3.390 (4.66)	4.482 (5.77)	3.879 (5.13)
ML	XGBoost / Cfb-7 (+m)	3.185 (4.42)	4.050 (5.28)	3.461 (4.65)
ML	XGBoost / Diverse-4	0.813 (1.23)	1.174 (1.68)	0.983 (1.44)
ML	XGBoost / Diverse-4 (+m)	0.810 (1.22)	1.148 (1.63)	0.988 (1.44)
ML	XGBoost / Diverse-7	0.797 (1.20)	1.159 (1.65)	0.981 (1.43)

Table 6. Task 3 Climate-diverse source-set transfer. Bold = overall best, underline = overall 2nd.

Figure 7. Directed cross-city transfer performance and UHI-regime coverage in the seven-city base setting. Directed transfer heatmaps for the seven-city base climate transfer experiment.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Data	Domain	Spatial	Number of	Temporal	# Temporal	Missing
Format	Type	Region	Grids/ Stations	Range	Step	Rate
Grid	Air Temperature UHI a	Berlin, DE	2,817	2015/01/01 – 2025/12/31	96,432	0.00%
		Hamburg, DE	2,216	2015/01/01 – 2025/12/31	96,432	0.00%
		Munich, DE	4,203	2015/01/01 – 2025/12/31	96,432	0.00%
		Cologne, DE	1,148	2015/01/01 – 2025/12/31	96,432	0.00%
		Dortmund, DE	864	2015/01/01 – 2025/12/31	96,432	0.00%

Appendix

Table 7. Full Dataset Summary for UHI-Bench

Figure 10. Relative distributions of Köppen–Geiger climate classification indicators for additional representative cities. The red markers show each city, and the shaded distributions provide benchmark-wide context. Appendix continuation of the climate-coverage figure, showing relative distributions of climate classification indicators for Berlin, Munich, Dortmund, Dusseldorf, Frankfurt, Stuttgart, Bucharest, Sao Paulo, and Cairo.

Feature	Aggregation Method	Formula	Output Unit
NDVI	Area-weighted mean (NDVI $>$ 0)	$\displaystyle\text{NDVI}_{i}=\frac{\sum_{p_{n}\in P_{i}^{+}}\text{NDVI}_{p_{n}}\cdot A_{p_{n}\cap i}}{\sum_{p_{n}\in P_{i}^{+}}A_{p_{n}\cap i}}$ , where $P_{i}^{+}=\{p_{n}\mid\text{NDVI}_{p_{n}}>0\}$ . If $P_{i}^{+}=\emptyset$ , $\text{NDVI}_{i}=0$ .	–
BCR	Area ratio	$\displaystyle\text{BCR}_{i}=\frac{\sum_{j=1}^{M}A_{b_{j}\cap i}}{A_{\text{grid},i}}$	–
Building Height	Area-weighted mean	$\displaystyle H_{i}=\frac{\sum_{j=1}^{M}h_{j}\cdot A_{b_{j}\cap i}}{\sum_{j=1}^{M}A_{b_{j}\cap i}}$	m
Water Surface Ratio	Area ratio	$\displaystyle W_{i}=\frac{\sum_{j=1}^{M}A_{w_{j}\cap i}}{A_{\text{grid},i}}$	–
Waterbody Proximity	Minimum distance	$\displaystyle D_{\text{water},i}=\min_{j}\{d(P_{i},W_{j})\}$	m
Road Density	Length density	$\displaystyle\text{RD}_{i}=\frac{\sum_{j=1}^{M}L_{r_{j}\cap i}}{A_{\text{grid},i}}\times 1000$	km/km 2

Appendix

Table 8. Aggregation formulas for the static grid features used in UHI-Bench.

Feature	Definition
Building coverage ratio (BCR)	the proportion of grid area covered by building footprints.
Mean building height	the area-weighted average height of all buildings intersecting a grid cell, giving greater influence to larger structures.
Normalized Difference Vegetation Index (NDVI)	seasonal mean greenness calculated from Sentinel-2 imagery and aggregated using area-weighted averaging across all intersecting pixels. MODIS LST is used separately for cross-validation of the downscaled LST product (Section 3 ) and is not the source of the NDVI feature.
Water coverage ratio	the proportion of grid area covered by water surface polygons, calculated as the intersection area between waterbody polygons and the grid cell divided by the total grid area.
Distance to nearest waterbody	Euclidean distance from the grid centroid to the nearest lake/pond (and coastal) water polygon.
POI density	number of Points of Interest per square kilometer, used as a proxy for human activity intensity and land-use mix.

Appendix

Table 9. Static urban morphology features used in UHI-Bench.

Table 10. Meteorological drivers (ERA5-Land) used in UHI-Bench. All six variables are obtained from ERA5-Land and resampled from their native grid to the standardized 1 km city grid (Section 3.1 ).

Group	Baselines	Main role	Input configurations
Statistical / geostatistical	Persistence, climatology, percentile rule, seasonal naive, inverse-distance weighting, ordinary kriging, universal kriging, regression kriging	Simple temporal and spatial references for forecasting, event detection, and imputation	base; +s where applicable
Classical machine learning	RF, XGBoost, Isolation Forest, One-Class SVM, linear models, regularized regressions	Covariate-aware imputation, detection, attribution, forecasting, and transfer analysis	base, +m, +s, +s+m when supported
Deep spatiotemporal models	SAITS, GNN/IGNNK-style models, GraphWaveNet, LSTM, PatchTST-style models, DLinear	Temporal, spatial, and graph-based inductive biases for imputation and forecasting	base, +m, +s, +s+m depending on model interface
Time-series foundation models	TimesFM, Chronos-2, MOIRAI-2, ChronosAdapter	Zero-shot or low-adaptation sequence modeling under missingness, sparse observations, forecasting, and cross-city transfer	TimesFM: base only; Chronos-2, MOIRAI-2, and ChronosAdapter: base, +m, +s, +s+m where supported

Appendix

Table 13. Baseline groups used in UHI-Bench.

Task	Evaluation protocol	Cities / source-target sets	Metrics
Analysis 1a: Cross-source consistency	Quantify LST-UHI/AirT-UHI consistency via temporal, spatial-pattern, and lagged correlation, plus cross-source imputation gains.	16 dual-source cities	Pearson ( $r$ ); $\Delta$ MAE
Analysis 1b: Heat-risk timing diagnostics	Identify source-specific extreme periods in city-mean hourly UHI using training-period P95 thresholds (min. 3 consecutive hours).	Berlin, Munich, Hamburg, Cairo, Lagos, Johannesburg, Riyadh	Local-hour share; night fraction; peak hour
Task 1: Extreme-event detection	Detect source-specific extreme events under base, +m, +s, and +s+m inputs.	Berlin, Munich, Hamburg, Cairo, Lagos, Johannesburg, Riyadh	F1 (main); FAR
Task 2a: LST cloud-gap imputation	Reconstruct LST-UHI using near-clear references and real cloud masks; errors computed over cloud-covered pixels across four coverage bins (0–25/25–50/50–75/ $>$ 75%). Evaluate base, +m, +s, +s+m where supported.	Cairo, Bucharest, Lagos	MAE (main); RMSE (supp.)
Task 2b: AirT sparse imputation	Reconstruct AirT-UHI from simulated station-like masks under the same four missing-rate bins. Evaluate base, +m, +s, +s+m where supported.	Lagos, Cologne, Riyadh, Munich, Cairo	MAE (main); RMSE (supp.)
Task 2c: Dual-source UHI forecasting	Forecast both UHI sources at 1/6/12/24/48/96 h horizons, comparing base, +m, +s, and +s+m inputs.	Munich, Berlin, Cologne, Lagos, Johannesburg, Riyadh	MAE (main); RMSE (supp.)

Appendix

Table 14. Complete task evaluation protocols, city coverage, and metrics used in the benchmark experiments.

		F1
		Berlin (Cfb)		Munich (Cfb)		Hamburg (Cfb)		Cairo (BWh)		Lagos (Aw)		Johannesburg (Cwb)		Riyadh (BWh)
Group	Model	AirT-UHI	LST-UHI	AirT-UHI	LST-UHI	AirT-UHI	LST-UHI	AirT-UHI	LST-UHI	AirT-UHI	LST-UHI	AirT-UHI	LST-UHI	AirT-UHI	LST-UHI
Stat/Geo	Percentile	0.894	0.197	0.883	0.020	0.891	0.054	0.774	0.511	0.654	0.248	0.760	0.621	0.786	0.329
Stat/Geo	SeasonalNaive	0.548	0.182	0.538	0.019	0.575	0.049	0.406	0.224	0.371	0.249	0.508	0.540	0.462	0.187
ML	RF	0.956	0.322	0.949	0.017	0.951	0.032	0.850	0.579	0.740	0.378	0.739	0.663	0.870	0.481
ML	RF+m	0.941	0.313	0.941	0.000	0.932	0.127	0.869	0.581	0.752	0.396	0.753	0.659	0.872	0.506

Appendix

Table 15. Complete Task 1 extreme-event detection results (F1, higher is better) for Berlin, Munich, Hamburg, Cairo, Lagos, Johannesburg, and Riyadh. The table adds the +s and +s+m F1 rows for the classical, unsupervised, and LSTM baselines; suffixes indicate input configuration.

		MAE
		Cairo				Bucharest				Lagos
Group	Model	0–25%	25–50%	50–75%	$>$ 75%	0–25%	25–50%	50–75%	$>$ 75%	0–25%	25–50%	50–75%	$>$ 75%
Stat/Geo	UnivKrig	1.318	1.435	1.829	2.343	1.283	1.539	1.650	2.101	1.616	1.626	1.919	2.185
Stat/Geo	IDW_static	1.405	1.568	2.082	2.258	0.989	1.379	1.285	1.440	1.500	1.656	2.005	2.521
ML	RF	1.639	1.682	2.390	2.527	0.964	1.403	1.344	1.514	1.554	1.779	2.079	2.511
ML	XGBoost	1.691	1.710	2.227	2.588	1.006	1.391	1.330	1.503	1.571	1.805	2.133	2.579

Appendix

Table 16. Additional Task 2a LST cloud-gap imputation results for baselines omitted from the compact main table. Each cell reports MAE. Method suffixes indicate input configuration. UnivKrig = Universal Kriging; GWNet = GraphWaveNet.

		MAE, by station missing-% bin
		Lagos (Aw)				Cologne (Cfb)				Riyadh (BWh)
Group	Method	0–25%	25–50%	50–75%	$>$ 75%	0–25%	25–50%	50–75%	$>$ 75%	0–25%	25–50%	50–75%	$>$ 75%
Stat/Geo	IDW	0.040	0.043	0.048	0.059	0.0992	0.1055	0.1141	0.1370	0.0585	0.0617	0.0671	0.0817
Stat/Geo	OrdKrig	0.055	0.064	0.081	0.125	0.1772	0.1975	0.2266	0.2559	0.0708	0.0794	0.0958	0.1525
Stat/Geo	UnivKrig	0.069	0.072	0.071	0.081	0.1288	0.1313	0.1399	0.1729	0.1042	0.1060	0.1051	0.1126
Stat/Geo	RegKrig+s	0.037	0.038	0.037	0.042	0.1020	0.1068	0.1112	0.1260	0.0666	0.0668	0.0638	0.0719

Appendix

Table 17. Continuation of Table 4 : Task 2b AirT station-sparse imputation MAE for additional Stat/Geo, ML covariate, and FM baselines. Each cell reports MAE by station-missing bin for Lagos, Cologne, and Riyadh. OrdKrig/RegKrig/UnivKrig = Ordinary/Regression/Universal Kriging.

Metric			MAE, AirT-UHI/LST-UHI
City	Group	Method	1h	6h	12h	24h	48h	96h	Avg
Munich	Stat/Geo	Persistence	0.006 /0.547	0.027/1.089	0.120/1.298	0.080/0.768	0.100/0.896	0.109/0.910	0.074/0.918
Munich	Stat/Geo	Climatology	0.096/0.972	0.085/0.940	0.016 /0.932	0.096/0.978	0.096/0.986	0.096/0.961	0.081/0.962
Munich	ML	XGBoost	0.011/0.499	0.042/0.792	0.056/0.642	0.052/0.606	0.061/0.621	0.064/0.666	0.048/0.637
Munich	ML	XGBoost+m	0.009 /0.474	0.025 /0.668	0.033/0.616	0.039 /0.628	0.045 /0.658	0.049 /0.655	0.033 /0.617
Munich	ML	XGBoost+s	0.010/0.498	0.034/0.780	0.045/0.636	0.049/0.605	0.056/ 0.612	0.061/0.672	0.042/0.634

Appendix

Table 19. Complete Task 2c forecasting results for Munich, Berlin, Cologne, Lagos, Johannesburg, and Riyadh. Each metric cell reports MAE, AirT-UHI/LST-UHI; Avg is the mean over the six displayed horizons.

Figure 11. Task 2d per-city signed top-3 driver rankings for AirT-UHI and LST-UHI in one view. Each city panel stacks the XGBoost top-3 meteorological drivers and top-3 static features for both targets; green bars denote AirT-UHI and orange bars denote LST-UHI. Bar direction is the raw high-minus-low response sign; bar length is normalized separately within each target’s meteorological and static groups, so that within-city ranking patterns remain visible. Sixteen small horizontal bar charts showing each city's signed top-three meteorological and static driver rankings for both AirT-UHI and LST-UHI.

Explore similar work

Apr 24, 2026

From physical surfaces to human-centric heat stress: LST and UTCI heat mapping reveals nonlinear effects of urban morphology

Heat exposure connects the built environment and public health, directly shaping the livability and sustainability of urban areas. Understanding the spatial heterogeneity of heat exposure and its drivers is vital for climate-adaptive urban planning. However, most planning-oriented studies rely on land surface temperature (LST), and whether LST adequately represents human heat exposure and how it differs from physiologically relevant heat stress remains insufficiently examined. Here, using Landsat-retrieved 30-m LST and GPU-accelerated 1-m universal thermal climate index (UTCI) in Singapore, this study establishes a comprehensive "Modeling-Comparing-Assessing" framework to systematically evaluate the spatial and mechanistic differences between these two metrics. We further investigate their pronounced non-stationary and threshold-based relationships with urban factors using a novel geographically weighted XGBoost (GW-XGBoost) and generalized additive model (GAM) workflow. Our results reveal substantial differences in the spatial patterns of LST and UTCI, along with marked spatial heterogeneity in how 2D and 3D urban factors impact these thermal metrics, as demonstrated by explainable GW-XGBoost models (test R2 = 0.855 for LST and 0.905 for UTCI). Crucially, spatially explicit SHAP shows that sky view factor plays a central role in explaining UTCI variability but exhibits a comparatively marginal independent contribution to LST, indicating that LST inadequately captures shading-driven and radiative processes governing actual human heat stress. Moreover, SHAP-GAM analysis indicates that higher albedo is associated with increased UTCI. These findings provide model-informed planning implications for integrating physiologically relevant thermal indices to support targeted heat risk management and human-centric urban planning.

Yuan Wang, Shengao Yi, Xiaojiang Li +4

Department of Architecture, National University of Singapore, Singapore 117566, Singapore · Department of City and Regional Planning, University of Pennsylvania, Philadelphia, PA 19104, USA · Urban Analytics Subject Group, Urban Studies & Social Policy Division, University of Glasgow +3

Urban Planning

Ask PDF

Jun 10, 2026

Urban Heat MiniCubes: An AI-Ready dataset for urban heat research

Urban heat is amplified by impermeable surfaces and heterogeneous built environments, yet street-level variability remains difficult to quantify because multi-sensor observations are rarely available in consistent, analysis-ready form at the necessary spatiotemporal scales. We present "Urban Heat MiniCubes," a publicly available, FAIR-oriented dataset designed for machine learning applications in urban heat research. The dataset provides harmonized 90 x 90 km gridded data cubes for 48 cities in the Western Hemisphere spanning 2022-2023, with variables reprojected and collocated to a common grid to reduce preprocessing (e.g., reprojection, resampling, and spatiotemporal alignment). Urban Heat MiniCubes includes two complementary modalities: (i) higher-spatial-resolution, lower-frequency observations from Landsat 8/9 (e.g., surface reflectances) and Sentinel-1 (e.g., synthetic aperture radar backscatter), and (ii) higher-temporal-frequency, coarser observations from GOES-R (e.g., longwave infrared brightness temperatures) and a microwave land surface temperature product. We document variables and metadata and provide technical assessment using inter-variable analyses and autoencoder-based reconstruction-error summaries across pixel classes (e.g., water and cloud). Potential use cases and limitations are also discussed.

Jonathan Starfeldt, Maria J. Molina, Alexander Kerr +3

Department of Atmospheric and Oceanic Science, University of Maryland, College Park, MD, USA. · Department of Computer Science, University of Maryland, College Park, MD, USA. · NASA Goddard Space Flight Center, Greenbelt, MD, USA. +1

Remote Sensing

Ask PDF

May 19, 2026

ShadeBench: A Benchmark Dataset for Building Shade Simulation in Sustainable Society

Urban heat exposure is becoming an increasingly critical challenge due to the intensifying urban heat island effect. Fine-grained shade patterns, especially those induced by urban buildings, strongly influence pedestrians' thermal exposure and outdoor activity planning. However, accurately modeling and analyzing urban shade at scale remains difficult because of the lack of large-scale datasets and systematic evaluation frameworks. To address this challenge, we present ShadeBench, a comprehensive dataset and benchmark for urban shade understanding. ShadeBench contains geographically diverse urban scenes with temporally varying simulated shade maps and textual descriptions, together with aligned satellite imagery, building skeleton representations, and 3D building meshes. Built upon this multimodal dataset, ShadeBench supports a range of downstream tasks, including shade generation, shade segmentation, and 3D building reconstruction. We further establish standardized evaluation protocols and baseline methods for these tasks. By enabling scalable and fine-grained shade analysis, ShadeBench provides a foundation for data-driven urban climate research and supports future studies in heat-resilient urban planning and decision-making. The code and dataset are publicly available at https://darl-genai.github.io/shadebench/.

Longchao Da, Mithun Shivakoti, Xiangrui Liu +3

School of Computing and Augmented Intelligence, Arizona State University · Global Futures Laboratory, Arizona State University

Benchmark Construction Synthetic Benchmark Generation

Ask PDF

		F1 (AirT-UHI\|LST-UHI)
Group	Model	Lagos (Aw)	Johannesburg (Cwb)	Cologne (Cfb)	Riyadh (BWh)
Stat/Geo	Percentile	0.654\|0.248	0.760\|0.621	0.875\|0.268	0.786\|0.329
Stat/Geo	SeasonalNaive	0.371\|0.249	0.508\|0.540	0.496\|0.231	0.462\|0.187
ML	RF	0.740\|0.378	0.739\|0.663	0.939 \|0.261	0.870\|0.481
ML	XGBoost	0.726\|0.361	0.713\|0.633	0.938\|0.226	0.876 \|0.460
ML	RF+m	0.752\|0.396	0.753\|0.659	0.923\|0.294	0.872\| 0.506