TLDR: A study fine-tuned a geospatial foundation model to predict urban land surface temperatures and simulate urban heat island effects under future climate and land cover scenarios. The model achieved high accuracy (errors below 1.74°C) and demonstrated strong extrapolation capabilities (up to 3.62°C beyond training data), effectively generalizing to extreme heat. It successfully simulated temperature changes from land cover modifications, highlighting the potential of AI for urban climate planning and mitigation, though further refinement is needed for complex spectral index interpretation.
As our cities continue to grow and climate change intensifies, a significant environmental challenge known as the urban heat island (UHI) effect is becoming more prevalent and severe. This phenomenon causes urban areas to be significantly warmer than their surrounding rural landscapes, leading to increased energy consumption, health issues, and declining air and water quality. Addressing this requires precise and timely temperature data, which traditional predictive methods often struggle to provide, especially in areas with limited data infrastructure.
A new approach is emerging through the use of geospatial foundation models (GFMs). These advanced AI models are trained on vast amounts of global, unstructured data, allowing them to generalize effectively across different spatial resolutions and geographic regions with minimal fine-tuning. This makes them a promising alternative for predicting temperatures where conventional methods fall short.
A recent study focused on fine-tuning a geospatial foundation model to predict urban land surface temperatures (LST) under various future climate scenarios. The research also explored how these temperatures respond to changes in land cover, specifically through simulated vegetation strategies. The chosen study area was Brasov, Romania, a city not included in the model’s initial training data, offering a challenging environment for UHI analysis due to its unique urban and environmental characteristics, including its location within the Carpathian Mountains.
The methodology involved extending an existing GFM to forecast UHI impacts under future Representative Concentration Pathway (RCP) scenarios and to simulate mitigation strategies. This was achieved through pixel-wise modifications and adjustments to spectral indices like the Normalized Difference Vegetation Index (NDVI), Normalized Difference Built-up Index (NDBI), and Normalized Difference Water Index (NDWI). These indices help characterize land cover types and their thermal effects, allowing researchers to simulate different urban designs, such as inserting forest pixels or urban pixels into satellite imagery.
The model was fine-tuned using a combination of open-access geospatial datasets from 2013 to 2023, including Landsat 8 satellite imagery, ERA5-Land reanalysis data, EURO-CORDEX climate model outputs, and ESRI land use/land cover (LULC) maps. To test the model’s ability to extrapolate to extreme heat conditions, the dataset was split by temperature, with the highest 10% of values reserved for testing.
The evaluation revealed impressive results. The fine-tuned model achieved pixel-wise downscaling errors below 1.74°C and accurately aligned with ground truth patterns. Crucially, it demonstrated an extrapolation capacity of up to 3.62°C, meaning it could successfully predict temperatures beyond its training range, effectively generalizing to potential future warming conditions in Brasov. This highlights the effectiveness of transfer learning with pretrained foundation models.
When analyzing performance across different land cover types, the model showed strong results in both vegetated areas (like “Trees” with an MAE of 1.73°C) and urban environments (“Built Area” with an MAE of 1.98°C). This indicates its relevance for urban climate modeling across dominant land cover types. Projections for 2030, 2050, and 2100, based on EURO-CORDEX data, clearly linked rising global temperatures to the expansion and intensification of UHI hotspots, underscoring the value of such forecasting for urban adaptation planning.
The model also proved capable of simulating landscape modification effects. Inserting forest pixels into a simulated area reduced predicted land surface temperatures, while inserting urban pixels increased them. This suggests the GFM can capture the spatial influence of land cover on surrounding microclimate conditions. However, the study noted that the model’s responsiveness to complex spectral indices like NDBI and NDWI was inconsistent with known physical relationships, suggesting it might rely more on RGB bands than other spectral channels for certain predictions.
Also Read:
- StefaLand: A New Geoscience Foundation Model for Dynamic Land-Surface Predictions
- Mapping Earth’s Clouds in 3D: A New Model for Satellite Data
In conclusion, this study demonstrates the significant potential of fine-tuned geospatial foundation models for predicting urban heat island effects under future climate scenarios and evolving urban landscapes. While GFMs offer strong generalization, further enhancements are needed to improve their predictive accuracy compared to highly localized deep learning models and to enhance their interpretability and physical realism. Future work may involve exploring alternative architectures, domain-specific constraints, and multimodal inputs. For more details, you can read the full research paper here.
