A Graph Neural Network Approach for Localized and High-Resolution Temperature Forecasting

AI summary

Overview

• Research area: Machine learning for weather and climate — specifically graph neural networks applied to localized, high-resolution surface temperature forecasting.
• Technical level: Intermediate. The paper assumes familiarity with neural network architectures (graph convolutional networks, gated recurrent units), representation learning, and standard forecasting evaluation metrics, but it is written in an accessible style with a strong equity framing.
• Scope (one sentence): The paper develops and evaluates a hybrid graph convolutional network plus gated recurrent unit (GCN–GRU) model that forecasts 2-meter air temperature 1–48 hours ahead at 2.5 km, hourly resolution over three nested regions of Southwestern Ontario, Canada, using NOAA URMA analysis data.

What This Paper Is About

Operational weather forecasts typically run at 10–30 km resolution, which smooths over urban heat islands and neighborhood-level hot spots, so the communities most vulnerable to heat — low-income, racialized, and Global South populations — often receive warnings that underestimate their actual risk. The paper's goal is to test whether a small, region-specific graph neural network trained on high-resolution gridded data can produce accurate, community-scale temperature forecasts that could eventually be post-processed into heatwave warnings and transferred to data-limited regions. The authors frame this as both a technical problem and an equity issue.

Key Contributions

1. A GCN–GRU forecasting framework at 2.5 km, hourly resolution that predicts 2-meter air temperature at seven horizons (1, 6, 12, 18, 24, 36, and 48 hours) across three nested bounding-box regions in Southwestern Ontario, each grid point modeled as a graph node with meteorological features.
2. A systematic scaling study across region sizes. The authors define Region A (approximately 44 km by 33 km), Region B (approximately 111 km by 163 km), and Region C (approximately 333 km by 243 km), and show that accuracy improves monotonically with domain size.
3. A language-model embedding pathway. Observations are serialized as short natural-language paragraphs and encoded with ClimateBERT into 768-dimensional vectors, then reduced with Principal Component Analysis and used as node features — a route to standardizing heterogeneous or missing inputs for data-limited settings.
4. A compute-versus-skill trade-off demonstration. A 6-hour sampling variant of the Region C model recovers most of the hourly model's accuracy at substantially lower compute, supporting the case for lightweight, locally deployable models.

Main Findings

• Larger spatial domain, better accuracy. Region A reached mean MAE 2.55 °C, MAE@48h 3.78 °C, and RMSE@48h 4.84 °C; Region B reached 2.48 °C, 3.73 °C, and 4.84 °C; Region C reached 1.93 °C, 2.93 °C, and 3.90 °C. The abstract reports the headline result as a mean MAE of 1.93 °C across 1–48 h forecasts and MAE@48h of 2.93 °C on the largest region, evaluated with 24-hour input windows.
• Coarser temporal sampling costs little. The Region C 6-hour model reached mean MAE 2.39 °C, MAE@48h 3.15 °C, and RMSE@48h 4.16 °C, versus 1.93 °C, 2.93 °C, and 3.90 °C for the hourly model — degradations of +0.46, +0.22, and +0.26 respectively. Training the full-resolution Region C model strained memory; the hourly Region C run used 2 NVIDIA H100 SXM (80GB) GPUs for 5 h 40 min, while the 6-hour run used 2 NVIDIA L40S (48GB) GPUs for 5 h 45 min.
• Random-weight control confirms learned signal. On Region A, a control model initialized with random weights performed far worse (mean MAE 9.11 °C, MAE@48h 8.89 °C, RMSE@48h 10.49 °C) than either the tabular baseline or the embedding model.
• Embedding performance is reported ambiguously. The Region A embedding model (ClimateBERT + PCA) scored mean MAE 3.34 °C, MAE@48h 4.34 °C, and RMSE@48h 5.54 °C against the tabular baseline's 2.55 °C, 3.78 °C, and 4.84 °C. The paper describes this as "a modest decrease in mean MAE relative to the tabular baseline," even though the reported embedding MAE is numerically higher than the baseline's; the authors present the embedding route primarily as a standardization mechanism for non-standard or missing data rather than an accuracy win.
• Compute and hardware vary by region. Region A trained in 35 min on 2 NVIDIA L40S GPUs; Region B in 42 min on 2 NVIDIA L40S GPUs; Region C in 5 h 40 min on 2 NVIDIA H100 SXM GPUs.
• Evaluation is limited. All results come from held-out test splits of URMA data. Benchmark systems such as GraphCast and FourCastNet operate at much coarser scales (roughly 10–50 km, 3–6 hour time steps), so no matched-resolution comparison is available.

Methodology in Plain English

The authors start from NOAA's Unrestricted Mesoscale Analysis (URMA) dataset, which provides hourly, 2.5 km gridded surface variables: 2-meter air temperature, 2-meter dewpoint, the u and v components of 10-meter wind, surface pressure, and orography (elevation). They carve out three nested rectangular regions over Southwestern Ontario that collectively include urban, farmland, forest, and water land types, chosen after discussions with climate NGOs in the Global South.

Each grid point becomes a node in a graph, carrying the meteorological features for that location and time. Nodes are connected to neighbors within a distance threshold (8 km for Regions A and B, 4 km for Region C). Graph convolution layers let each node incorporate information from its neighbors, modeling spatial interactions; gated recurrent units then model how conditions evolve over time. The model is trained with mean squared error to predict temperature at 1, 6, 12, 18, 24, 36, and 48 hours ahead.

Data were standardized using training-set statistics, with missing or invalid entries filled by spatial interpolation (the mean of nearby grid points). The target temperature was normalized during training but all reported errors are in degrees Celsius. Splits were region-specific: Regions A and B used manual splits (2022 train, 2023 validation, 2024-to-end test), while Regions C used a 70%/15%/15% ratio split, with Region C covering 18,396 / 3,942 / 3,943 timestamps and the 6-hour variant covering 8,170 / 1,751 / 1,751. Hyperparameters were selected via grid search — for example Region A used a learning rate of 0.0001, a 48-hour input window, batch size 16, hidden dimension 32, and an 8 km connectivity distance.

In a parallel experiment, each Region A observation was converted into a short natural-language description of its variables, encoded with ClimateBERT into a 768-dimensional vector, reduced with PCA (fit on the training set), and fed into the same GCN–GRU pipeline as node features. This tests whether language-model representations can standardize messy, incomplete inputs.

Why This Matters

The paper argues that coarse global models systematically miss the neighborhood-scale heat extremes that matter most to vulnerable populations, and that lightweight, region-specific models can deliver outsized practical value relative to models requiring massive compute. It situates itself in the climate AI discussion by showing that small GNNs trained on a single region can reach sub-3 °C errors at 48 hours, and that the same framework could be adapted for under-monitored contexts.

Real-world applications the authors identify or imply:

• Heatwave early-warning systems that operate at community rather than regional scale, since the forecast temperature is a low-level signal that can be post-processed against jurisdiction-specific heatwave definitions (consecutive hot days, absolute thresholds, or percentile criteria).
• Targeted allocation of scarce resources such as cooling centers, medical services, and energy, particularly for Global South governments with limited budgets.
• Municipal planning and public health interventions that need neighborhood-level rather than city-average temperature projections.
• Standardized data ingestion for data-poor regions, via the ClimateBERT plus PCA embedding pathway, which is designed to accommodate missing or unstandardized variables.

Industry relevance spans climate and weather technology, energy utilities managing demand spikes, public health agencies, insurers assessing heat risk, and any organization that needs localized forecasts deployable on modest hardware — the paper emphasizes that even the largest region trained in under six hours on two GPUs.

Future Directions

1. Matched-resolution baselines. The authors explicitly call for comparisons against state-of-the-art systems re-run at comparable resolution and time step, since the current 10–50 km and 3–6 hour mismatch makes fair benchmarking impossible.
2. Transfer learning to data-sparse regions. The central stated ambition is to test whether models trained in data-rich Southwestern Ontario can be adapted to under-monitored regions of the Global South and to different climate regimes, with minimal fine-tuning as local data becomes available.
3. Expanding the predicted variables. The authors plan to forecast humidity, wind, and related variables beyond temperature to build a composite index for more reliable extreme-event detection.
4. Operational integration and other extremes. They propose integrating with operational dashboards for timely warnings and extending the framework to wildfires, floods, and droughts. A related open question is whether the embedding-based input standardization can close the accuracy gap with the tabular baseline once more local data is available.

Target Audience

This paper is most useful to machine learning researchers working on geospatial and spatiotemporal forecasting, climate AI practitioners interested in graph neural networks and representation learning for Earth observation data, and applied researchers or policymakers focused on heat-health early warning and climate equity. It is also relevant to engineers building lightweight, locally deployable forecasting systems who want a concrete demonstration of resolution, compute, and accuracy trade-offs at the 2.5 km and hourly scale. Readers seeking rigorous cross-model benchmarking at matched resolution, or validated transfer results for the Global South, will not find them here — the paper presents the Southwestern Ontario case as a foundation rather than a completed transfer study.

Authors’ abstract

Heatwaves are intensifying worldwide and are among the deadliest weather disasters. The burden falls disproportionately on marginalized populations and the Global South, where under-resourced health systems, exposure to urban heat islands, and the lack of adaptive infrastructure amplify risks. Yet current numerical weather prediction models often fail to capture micro-scale extremes, leaving the most vulnerable excluded from timely early warnings. We present a Graph Neural Network framework for localized, high-resolution temperature forecasting. By leveraging spatial learning and efficient computation, our approach generates forecasts at multiple horizons, up to 48 hours. For Southwestern Ontario, Canada, the model captures temperature patterns with a mean MAE of 1.93$^{\circ}$C across 1-48h forecasts and MAE@48h of 2.93$^{\circ}$C, evaluated using 24h input windows on the largest region. While demonstrated here in a data-rich context, this work lays the foundation for transfer learning approaches that could enable localized, equitable forecasts in data-limited regions of the Global South.