Research map / Label-efficient and robust learning / Recognition and long-tail
Measuring Fine-Grained Urban Air Temperature with Satellite Imagery
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes UrbanHeat, which predicts fine-grained (0.25km) urban air temperature from multi-resolution satellite imagery using LST ranking loss and global climate data (CHELSA) regression as proxy labels.
Problem
Weather stations lack spatial coverage and resolution for fine-grained urban air temperature measurement, and existing alternatives either miss regional land-cover characteristics or require hard-to-obtain dynamic climate inputs.
Approach
Uses two ResNet-18 encoders on high- and low-resolution satellite images with multi-view cross-attention, plus latitude and elevation, trained with a differentiable Spearman-style ranking loss on grouped, filtered LST values and a regression loss to global climate data (CHELSA) so no station labels are needed.
Key results
- UrbanHeat achieves Spearman 0.625 in Hong Kong and 0.533 in Seoul, the best among compared methods (CHELSA 0.510 and 0.398).
- Pearson 0.958 in Hong Kong and MAE 0.469 in Hong Kong; in Seoul Pearson 0.579 and MAE 0.512.
- Removing the LGCD loss drops Hong Kong Pearson to 0.650 and MAE to 3.707, showing the scale loss is essential.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add LST-based ranking loss (ablation) | UrbanHeat · Hong Kong summer 2019 | Spearman ρs: 0.518 -> 0.625 | Helped |
| add LST-based ranking loss (ablation) | UrbanHeat · Seoul summer 2019 | Spearman ρs: 0.438 -> 0.533 | Helped |
| add GCD-based scaling loss (ablation) | UrbanHeat · Hong Kong summer 2019 | Spearman ρs: 0.492 -> 0.625 | Helped |
| add GCD-based scaling loss (ablation) | UrbanHeat · Seoul summer 2019 | Spearman ρs: 0.472 -> 0.533 | Helped |
| use cross-attention instead of concatenating image embeddings (ablation) | UrbanHeat · Hong Kong summer 2019 | Spearman ρs: 0.607 -> 0.625 | Helped |
| use cross-attention instead of concatenating image embeddings (ablation) | UrbanHeat · Seoul summer 2019 | Spearman ρs: 0.472 -> 0.533 | Helped |
| use both image resolutions instead of only the high-resolution image (ablation) | UrbanHeat · Hong Kong summer 2019 | Spearman ρs: 0.603 -> 0.625 | Helped |
| use both image resolutions instead of only the high-resolution image (ablation) | UrbanHeat · Seoul summer 2019 | Spearman ρs: 0.485 -> 0.533 | Helped |
| use both image resolutions instead of only the low-resolution image (ablation) | UrbanHeat · Hong Kong summer 2019 | Spearman ρs: 0.572 -> 0.625 | Helped |
| use both image resolutions instead of only the low-resolution image (ablation) | UrbanHeat · Seoul summer 2019 | Spearman ρs: 0.477 -> 0.533 | Helped |
| use UrbanHeat instead of CHELSA | UrbanHeat · Hong Kong summer 2019 | Spearman ρs: 0.510 -> 0.625 | Helped |
| use UrbanHeat instead of Top. Reg | UrbanHeat · Seoul summer 2019 | Spearman ρs: 0.449 -> 0.533 | Helped |
Surprising findings
- Raw LST values alone correlate poorly with station air temperature (Spearman 0.246 in Hong Kong, near zero in Seoul), and ERA-5 has the lowest correlation.
- Simple concatenation of the two image embeddings did not outperform a single-image model, while cross-attention did.
Limitations
- Evaluation uses a limited number of ground-truth weather stations (37 in Hong Kong, 62 in Seoul) for two cities and one summer.
- The model does not directly incorporate physical weather principles and relies on GCD to carry them indirectly.
- Relies on proxy labels (LST and CHELSA) rather than true air temperature for training.
Experimental setup
Datasets: Esri World Imagery SRTM 1 Arc-Second Global Landsat 8 Collection 2 LST CHELSA Hong Kong CSDI AWS data KMA Seoul station data NGII grid population data S-DoT dataset DIV2K
Benchmarks: Hong Kong AWS summer 2019 temperatures Seoul KMA summer 2019 temperatures
Metrics: Spearman correlation Pearson correlation MAE
Baselines: ERA-5 SRGAN Top. Reg ResNet Tile2Vec LST CHELSA Proxy Ens.
Builds on: ResNet Tile2Vec SRGAN Spearman correlation
Closest papers by meaning
- UrbanVLP: Multi-Granularity Vision-Language Pretraining for Urban Socioeconomic Indicator Prediction (AAAI 2025 · similarity 0.84)
- UrbanDataLayer: A Unified Data Pipeline for Urban Science (NeurIPS 2024 · similarity 0.83)
- Terra: A Multimodal Spatio-Temporal Dataset Spanning the Earth (NeurIPS 2024 · similarity 0.83)
- CityLens: Evaluating Large Vision-Language Models for Urban Socioeconomic Sensing (ICLR 2026 · similarity 0.82)
- EcoMapper: Generative Modeling for Climate-Aware Satellite Imagery (ICML 2025 · similarity 0.82)
- CityPulse: Fine-Grained Assessment of Urban Change with Street View Time Series (AAAI 2024 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.