Open Source Project for Predicting Urban Land Surface Temperature Using Graph Neural Networks

The open source project introduced in this article is urban-lst-prediction-gnn (link: https://github.com/LiboRom/urban-lst-prediction-gnn), released by Liborio Román Montes (LiboRom) on GitHub on June 4, 2026. Based on Graph Convolutional Networks (GCN), this project builds a spatiotemporal prediction system for urban land surface temperature (LST). It integrates satellite remote sensing, meteorological data, and urban morphological indicators to provide an innovative solution for urban thermal environment monitoring. Its core goal is to improve the prediction accuracy and application value of issues related to the urban heat island effect.

With the acceleration of global urbanization, the urban heat island effect has become a key environmental issue affecting residents' quality of life and sustainable development. Accurate LST prediction is of great significance for urban planning, energy management, and public health. Traditional meteorological prediction methods struggle to capture the complex microclimate characteristics inside cities, while machine learning technologies provide new ideas to solve this problem. This project is the core result of Liborio Román Montes' master's thesis at the University of Seville (Master's in Artificial Intelligence). It innovatively applies Graph Neural Networks (GNN) to spatiotemporal LST prediction, constructing a high-precision prediction system through multi-source data integration.

Multi-source Data Fusion

The project integrates multiple heterogeneous data sources:

• Satellite Remote Sensing: Landsat 8/9 (provides LST, NDVI, etc., from Google Earth Engine), Sentinel-3 (used for downscaling experiments, from Copernicus Data Space Ecosystem);
• Meteorological Data: ERA5-Land reanalysis dataset (air temperature, humidity, solar radiation, etc., from ECMWF);
• Urban Morphology Data: DEM and urban morphological variables from CNIG, building indicators and geometric descriptors from OpenStreetMap.

Graph Structure Modeling

The study area (Seville urban area) is discretized into a regular spatial grid. Each node represents an urban location, and edges connect adjacent nodes, enabling GCN to effectively utilize spatial dependencies and capture the neighborhood effects of the urban thermal environment.

Core Model Architecture

The project implements and compares multiple models:

• GNN Models: GCN (spectral domain convolution to aggregate neighborhood information), GCN+GRU (combining spatiotemporal features), GraphSAGE (inductive reasoning);
• Traditional Deep Learning Baselines: LSTM, GRU;
• Machine Learning Baselines: XGBoost.

Evaluation Metrics

Model performance is evaluated using RMSE (Root Mean Square Error), MAE (Mean Absolute Error), R² (Coefficient of Determination), and MAPE (Mean Absolute Percentage Error).

Code Structure

• TFM_LIBORIO_v2.ipynb: Main notebook covering data acquisition and preprocessing, graph construction, Sentinel-3 downscaling, model implementation, and evaluation;
• training_automatization.ipynb: Automated experiment notebook supporting hyperparameter research, ablation experiments, multi-random seed evaluation, and model comparison.

Dependency Environment

Python 3.x, with main dependencies including PyTorch/TensorFlow, PyTorch Geometric/DGL (GNN libraries), Google Earth Engine API, XGBoost, NumPy/Pandas/Scikit-learn, etc.

External Service Integration

Requires access to Google Earth Engine, Copernicus Data Space Ecosystem, ERA5-Land archives (CDS API), and CNIG services to obtain data.

Academic Contributions

Innovatively introducing GNN into urban climate modeling, explicitly encoding the mutual influence between urban spatial units. Experiments show that GCN and its variants have significant advantages in capturing local thermal environment features.

Practical Applications

• Urban Planning: Identify high-risk heat island areas, guide green space layout and architectural design;
• Energy Management: Predict peak cooling demand, optimize power grid scheduling;
• Public Health: Early warning of extreme high-temperature events, protect vulnerable populations;
• Climate Change Research: Evaluate the response of urban thermal environments to climate change.

Methodological Insights

Demonstrate the potential of multi-source data fusion and interdisciplinary methods (remote sensing, meteorology, machine learning) in environmental science, opening up new directions for urban climate science.

Current Limitations

• Data Dependency: Reproduction requires access permissions to external services such as Google Earth Engine and Copernicus;
• Computing Resources: Large-scale GNN training requires high computing resources;
• Geographic Limitation: Only applied to Seville urban area; generalization to other climate zones needs verification.

Future Directions

• Introduce attention mechanisms (GAT) to enhance modeling of key spatial relationships;
• Explore advanced architectures like Spatiotemporal Graph Convolutional Networks (ST-GCN);
• Expand to more cities and build a general prediction framework;
• Combine physical constraints to develop a Physics-Informed Neural Network (PINN) version.

This project is a successful application of GNN in urban environmental science. Through innovative data fusion strategies and rigorous model comparisons, it not only provides a high-precision prediction tool but also lays a methodological foundation for subsequent research. As urban climate challenges intensify, this open source project will provide valuable technical resources for researchers and urban planners worldwide.