💻 computer science

HeatCast: A Benchmark for Neighborhood-Scale LST Forecasting across 124 U.S. Cities

The paper introduces HeatCast, a comprehensive benchmark for neighborhood-scale (30 m) monthly Land Surface Temperature forecasting across 124 U.S. cities that includes a large-scale dataset, standardized evaluation metrics, and demonstrates the superior performance of the Earthformer model over CNN+LSTM baselines.

Original authors: Jesus Guerrero, Isaac Corley, Leon Najafirad, Maryam Tabar, Paul Rad

Published 2026-08-11

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Jesus Guerrero, Isaac Corley, Leon Najafirad, Maryam Tabar, Paul Rad

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the Earth as a giant, glowing patchwork quilt. Some squares are made of dark asphalt and concrete, soaking up the sun like a black car in July. Others are made of green grass or blue water, which stay cool and fresh. This difference in temperature isn't just about comfort; it's a matter of life and death. When cities get too hot, it can make people sick or even cause them to die. Scientists call this "Land Surface Temperature" (LST), which is basically a fancy way of measuring how hot the ground is, rather than the air.

For a long time, scientists have tried to predict this heat, but their maps were often blurry. They looked at huge chunks of land, like viewing a city through a foggy window, missing the tiny details where the heat actually builds up. They also mostly looked at just one or two cities at a time, making it hard to compare if a new method was actually good or just lucky. To fix this, researchers needed a giant, clear, shared playground where everyone could test their prediction tools on the same tricky neighborhood-scale puzzles.

Enter HeatCast, a new project that acts like a massive, high-definition video game level for predicting urban heat. Instead of looking at whole cities, this team created a benchmark that zooms in on individual neighborhoods, block by block, across 124 different U.S. cities. They gathered a decade's worth of satellite photos, from 2013 to mid-2025, to create a dataset so detailed it can see the temperature difference between a hot parking lot and a cool park right next door.

The researchers didn't just dump the data; they built a strict "exam" for computer models to take. They set up a specific challenge: look at the last 12 months of weather and land data, and guess what the ground temperature will be next month. They tested two different types of AI "students" on this exam. One was a classic, reliable student (a CNN+LSTM), and the other was a newer, more advanced student called Earthformer.

The results were a bit surprising. The advanced Earthformer model did much better, predicting the heat with an average error of 7.74 K (Kelvin), while the classic model was off by 10.42 K. But the most interesting discovery wasn't just about which model was smarter; it was about what they used to think. The researchers found that the best predictions didn't come from just looking at past heat records. In fact, the Earthformer model got its lowest error rate of 7.72 K when it ignored the past heat data entirely and instead focused on eight other clues, like how much green vegetation was there, how much water was visible, and how shiny the ground was. It turns out that knowing what the ground looks like is often a better crystal ball for predicting how hot it will get than just knowing how hot it was yesterday.

This paper doesn't claim to have solved the heat problem forever, but it has handed the scientific community a powerful new tool. By releasing all their data, code, and the "answer keys" for their models, they've given everyone a fair way to build better heat-prediction tools. This could help city planners figure out exactly where to plant trees or build parks to cool down the hottest, most vulnerable neighborhoods, making cities safer and more comfortable for everyone.

Technical Summary: HeatCast – A Benchmark for Neighborhood-Scale LST Forecasting

Problem Statement

Land Surface Temperature (LST) is a critical metric for understanding urban heat, yet there is no widely adopted, open benchmark for forecasting LST at the neighborhood scale (30 m resolution). Prior research is limited by three main factors:

Geographic Scope: Most studies focus on only one to three cities.
Resolution: Existing datasets often use kilometer-scale products that smooth out critical intra-city thermal gradients (e.g., differences between impervious surfaces and vegetation).
Reproducibility: Many studies do not release data, code, or standardized evaluation protocols, making it difficult to compare model performance or task difficulty across studies.

Consequently, the community lacks a standardized way to evaluate models for next-month LST forecasting across diverse urban environments.

Methodology

Dataset Construction (HeatCast)

The authors constructed HeatCast, a benchmark dataset covering 124 U.S. cities with urban footprints larger than 90 mi².

Temporal Coverage: Mid-2013 (Landsat 8 commissioning) through June 2025.
Spatial Resolution: 30 m grid, derived from Landsat 8/9 Collection 2 Level-2 data.
Data Volume: Approximately 1.4 million 128×128 pixel patches (~150 GB).
Input Channels (9 per tile):
- Land Surface Temperature (LST).
- Digital Elevation Model (DEM).
- Surface Reflectance RGB (Bands 2–4).
- Spectral Indices: NDVI, NDWI, and NDBI.
- Broadband Albedo.
Labels: Each pixel is assigned a Local Climate Zone (LCZ) label (17 classes + unclassified) to enable stratified evaluation.
Preprocessing: Scenes are mosaicked, reprojected to a common UTM grid, and cloud-affected pixels are masked. Missing months within a sequence are handled via linear interpolation (up to two consecutive months) or excluded.

Benchmark Protocol

Task: Next-month LST forecasting. Given 12 consecutive months of 9-channel observations, predict the LST channel for the following month.
Temporal Split:
- Training: 2013–2021.
- Validation: 2022–2023.
- Test: January 2024–June 2025.
Metrics: Root Mean Square Error (RMSE) in Kelvin, reported as a mean over valid pixels. Evaluation is stratified by LCZ super-clusters (Compact Urban, Open Urban, Other Urban, Natural).

Baseline Models

Two architectures were implemented and trained on the temporal split:

CNN+LSTM: A two-layer convolutional encoder (64 hidden channels) followed by a 128-unit LSTM for temporal modeling and a linear decoder.
Earthformer: A transformer-based model using cuboid attention for spatio-temporal forecasting, configured with 256 base units, 4 attention heads, and a stack-conv downsampler.

Both models were trained using AdamW (learning rate $5 \times 10^{-4}$ ) with early stopping and data augmentation (rotations/flips).

Key Results

Baseline Performance

Earthformer achieved an aggregate test-set RMSE of 7.74 K.
CNN+LSTM achieved an aggregate test-set RMSE of 10.42 K.
LCZ Stratification: Earthformer outperformed CNN+LSTM in Open Urban, Other Urban, and Natural categories. However, CNN+LSTM performed better on Compact Urban pixels (8.62 K vs. 12.68 K), which constitute only 0.76% of the dataset.

Feature Ablation (Earthformer)

The authors investigated the predictive value of different input channel subsets:

Auxiliary Channels Only (No LST): Using the 8 non-LST channels (DEM, RGB, Indices, Albedo) yielded the lowest RMSE of 7.72 K.
All Channels: 7.74 K.
LST History Only: 8.15 K.
RGB Only: 8.68 K.

This indicates that for Earthformer, auxiliary variables provide a stronger predictive signal for next-month LST than historical LST alone. Conversely, for the CNN+LSTM, removing historical LST significantly degraded performance (RMSE increased from 10.42 K to 12.47 K).

Significance and Claims

The paper claims the following contributions and significance:

First Neighborhood-Scale Benchmark: HeatCast is the first shared benchmark for monthly LST forecasting at 30 m resolution across a large number of U.S. cities (124), addressing the gap left by prior kilometer-scale or single-city studies.
Reproducibility: The dataset, code, fixed temporal splits, and reference evaluation harness are released under the MIT license, allowing for direct comparison of future models.
Insight into Predictive Features: The ablation study reveals that non-LST channels (specifically auxiliary variables like albedo and spectral indices) can outperform historical LST data in transformer-based models for this specific forecasting task.
Foundation for Future Work: The benchmark supports research into class-balanced forecasting (particularly for rare compact urban classes), feature fusion, and the application of geospatial foundation models to thermal forecasting.

The authors note limitations, including the inherent uncertainty of Landsat LST products (1–2 K under clear skies), the focus on large cities with sufficient cloud-free imagery, and the fact that the benchmark evaluates future dates in known cities rather than transfer to unseen geographies. The dataset is available at https://doi.org/10.57967/hf/9889.