💻 computer science

Spatiotemporal downscaling and nowcasting of urban land surface temperatures with deep neural networks

This paper presents a deep learning framework that combines geostationary and polar-orbiting satellite data to generate high-resolution (1 km, 15-min) Land Surface Temperature fields via a U-Net downscaling model and subsequently forecasts intraday LST variations using a ConvLSTM nowcasting model, achieving high accuracy across major European cities.

Original authors: Solomiia Kurchaba, Angela Meyer

Published 2026-06-19

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Solomiia Kurchaba, Angela Meyer

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 trying to watch a live weather movie of a city, but you only have two types of cameras:

The "Fast" Camera: It takes a picture of the whole city every 15 minutes, but the image is blurry (like a low-resolution phone photo). It sees the general heat, but misses the details.
The "Sharp" Camera: It takes a super-clear, high-definition picture of the city, but it only flies over once or twice a day. It sees every hot sidewalk and cool park, but you have to wait a long time for the next shot.

The problem is that to understand how cities heat up during the day (like during a heatwave), you need both the speed of the first camera and the sharpness of the second.

This paper presents a clever solution using a "smart AI" that acts like a digital editor, combining these two cameras to create a high-definition movie that updates every 15 minutes.

Here is how they did it, broken down into two main steps:

Step 1: The "Super-Res" Filter (Downscaling)

First, the researchers trained a deep learning model (specifically a U-Net, which is like a digital artist that knows how to fill in missing details) to turn the blurry "Fast" camera images into sharp ones.

How it works: The AI looks at the blurry 3-kilometer-wide pixels from the European weather satellite (SEVIRI) and learns to guess what the 1-kilometer details should look like, based on the few sharp images it has seen from the other satellite (MODIS).
The Secret Ingredient: They also fed the AI the "Sun Angle" (how high the sun is in the sky). This helps the AI understand where the sun is hitting and where shadows will fall, making the temperature guesses much smarter.
The Result: They successfully created a map of surface temperatures across major European cities that is sharp (1 km detail) and updates every 15 minutes. It's like turning a pixelated sketch into a crisp photograph, and doing it continuously.

Step 2: The "Crystal Ball" (Nowcasting)

Once they had these sharp, 15-minute updates, they needed to predict what would happen in the next 15 to 75 minutes. This is called "nowcasting."

The Tool: They used a different type of AI called ConvLSTM. Think of this as a model that watches a short video clip of the city heating up and learns the "flow" of the heat. It understands that heat doesn't just jump around randomly; it moves and changes in patterns.
The Test: They tested this on three very different cities: Bucharest (hot and humid), Antwerp (temperate), and Berlin (transitional).
The Performance: The AI was much better at guessing the future temperature than two simple "lazy" methods:
- The "Stuck" Method: Assuming the temperature stays exactly the same as the last minute (Persistence).
- The "History Book" Method: Assuming the temperature will be the same as it was at this exact time over the last five days (Climatology).
- The AI beat both of them, predicting temperatures with very high accuracy (errors of less than 1 degree Celsius for short-term forecasts).

Why This Matters (According to the Paper)

The authors explain that this system is a game-changer for monitoring Urban Heat Islands (where cities get much hotter than the surrounding countryside).

Day vs. Night: The system works best at night. At night, the city cools down in a smooth, predictable way. During the day, the sun creates complex patterns of hot and cold spots (shadows, hot roofs, cool parks), which makes it harder for the AI to be perfect, though it still performs well.
No Extra Data Needed: The AI is smart enough to figure out the details just by looking at the temperature patterns and the sun's position. It doesn't need a separate database of building heights or tree types, which makes it easier to use in many different cities.

The Bottom Line

The researchers built a two-step AI pipeline:

Sharpen the blurry, fast satellite images to get a clear, 15-minute update of city temperatures.
Watch the pattern of those clear images to predict the next hour of temperature changes.

They proved this works across Europe, offering a tool that can help city planners and emergency services see and predict heat stress in real-time, specifically for the summer months when it matters most.

Technical Summary: Spatiotemporal Downscaling and Nowcasting of Urban Land Surface Temperatures with Deep Neural Networks

Problem Statement
Land Surface Temperature (LST) is a critical variable for monitoring urban heat islands, climate change, and energy demand. However, a fundamental trade-off exists in satellite-derived LST products: geostationary sensors (e.g., SEVIRI/Meteosat) offer high temporal resolution (15-minute intervals) but lower spatial resolution (~3 km), while polar-orbiting sensors (e.g., MODIS/Terra and Aqua) provide high spatial resolution (1 km) but only capture 2–4 overpasses per day. This gap hinders the ability to study sub-diurnal LST processes and perform intraday forecasting at high resolution. Existing downscaling studies often rely on classical machine learning with human-defined features or deep learning models trained on specific cities with limited data availability (e.g., using ECOSTRESS with long revisit times), making them difficult to scale across diverse urban environments. Furthermore, intraday forecasting of high-resolution LST fields remains a limited research area.

Methodology
The authors propose a two-stage deep learning framework to address these limitations:

Spatiotemporal Downscaling (U-Net):
- Objective: To map low-resolution SEVIRI LST fields (3 km, 15-min) to high-resolution MODIS-like LST fields (1 km).
- Input: SEVIRI LST fields augmented with Solar Zenith Angle (SZA) as an auxiliary predictor.
- Target: Co-located, high-resolution MODIS LST (Terra/Aqua) observations.
- Architecture: A U-Net convolutional encoder–decoder model. The contracting path captures contextual information, while the expansive path reconstructs high-resolution details. Skip connections preserve fine-scale spatial patterns.
- Data: Training data covers major European cities (population > 1 million) during the warm season (May 15–September 15) from 2004–2025. The dataset includes 52,938 data points, with strict quality filtering for clouds and viewing angles.
- Strategy: The model is trained to generalize across diverse climatic and morphological conditions without relying on static auxiliary datasets (e.g., land cover maps), assuming their effects are implicitly encoded in the coarse LST signal.
Intraday Nowcasting (ConvLSTM):
- Objective: To forecast high-resolution (1 km) LST fields for lead times of 15 to 75 minutes.
- Input: Time-series sequences of the downscaled LST fields generated by the U-Net model (three consecutive 15-min steps).
- Target: Future downscaled LST fields at the specified lead time.
- Architecture: A Convolutional LSTM (ConvLSTM) network, which combines recurrent temporal memory with convolutional operators to capture both spatial features and temporal dependencies simultaneously.
- Benchmarks: The model is compared against a Persistence benchmark (assuming no change) and a Climatological Rolling Median benchmark (median of the same time slot over the previous five days).

Key Contributions

First Pan-European Downscaling Model: The study presents the first deep learning model capable of generating 1 km resolution LST fields updated every 15 minutes across multiple major European cities, mapping 3-km SEVIRI data to 1-km MODIS targets.
Generalizability: Unlike previous city-specific approaches, this model is designed to generalize across diverse urban areas (population > 1 million) with low error and near-zero bias.
First Intraday Deep Learning Nowcaster: The authors introduce the first satellite-based deep learning nowcasting model for high-resolution LST fields, demonstrating its application in Bucharest, Antwerp, and Berlin with lead times up to 75 minutes.
Two-Step Pipeline: The framework successfully bridges continental-scale data generation (downscaling) with localized operational forecasting (nowcasting).

Results

Downscaling Performance: On a hold-out test set, the U-Net model achieved an $R^2$ of 0.97, a Root Mean Square Error (RMSE) of 1.92 °C, and a Mean Bias Error (MBE) of 0.01 °C. The model showed robust performance across different cities, months, and hours, with near-zero bias indicating no systematic over- or underestimation. Performance was slightly better at night than during the day.
Nowcasting Performance: The ConvLSTM models significantly outperformed both Persistence and Climatological Rolling Median benchmarks.
- RMSE: Ranged from 0.57 °C to 1.15 °C depending on the lead time (15 to 75 minutes) and city.
- Bias: MBE values were close to zero (ranging from -0.1 to 0.14 °C).
- Diurnal Variation: The model performed well throughout the day but faced slightly more challenge during transition periods (around 14:00 and 03:00–05:00) where temperature tendencies change sign.
Independent Validation: When validated against independent MODIS overpasses (not used in training), the pipeline showed strong agreement, particularly at night ( $R^2$ 0.82–0.97, RMSE 0.88–1.44 °C). Daytime performance was lower ( $R^2$ 0.44–0.89, RMSE 1.96–2.7 °C), attributed to higher spatial variability driven by solar radiation and surface heterogeneity.

Significance and Claims
The paper claims that this two-step approach provides a feasible and effective strategy for the diurnal monitoring of urban LST. The authors assert that their framework offers:

Operational Applicability: The model is directly applicable to operational satellite-based LST monitoring.
Practical Utility: The system supports near-real-time urban heat island monitoring, early warning for heat stress episodes, and improved city-scale heat-risk mapping.
Robustness: The approach demonstrates strong predictive skill and low systematic bias, outperforming standard data-driven benchmarks.
Scalability: By avoiding reliance on heterogeneous static auxiliary datasets, the method is designed to be transferable across different urban environments.

The authors note limitations, including the restriction to the warm season (May–September) and the sequential nature of the pipeline, where errors from the downscaling stage propagate to the nowcasting stage. They suggest that future work could explore the inclusion of additional dynamic and stationary auxiliary variables (e.g., NDVI, urban geometry) to further improve physical interpretability and performance, though this may introduce complexity in model transferability.

Step 1: The "Super-Res" Filter (Downscaling)

Step 2: The "Crystal Ball" (Nowcasting)

Why This Matters (According to the Paper)

The Bottom Line

Technical Summary: Spatiotemporal Downscaling and Nowcasting of Urban Land Surface Temperatures with Deep Neural Networks

More like this