Harnessing Multi-Source Data and Deep Learning for High-Resolution Land Surface Temperature Gap-Filling Supporting Climate Change Adaptation Activities
Addressing global warming and adapting to the impacts of climate change is a primary focus of climate change adaptation strategies at both European and national levels. Land surface temperature (LST) is a widely used proxy for investigating climate-change-induced phenomena, providing insights into t...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
MDPI AG
2025-01-01
|
| Series: | Remote Sensing |
| Subjects: | |
| Online Access: | https://www.mdpi.com/2072-4292/17/2/318 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Addressing global warming and adapting to the impacts of climate change is a primary focus of climate change adaptation strategies at both European and national levels. Land surface temperature (LST) is a widely used proxy for investigating climate-change-induced phenomena, providing insights into the surface radiative properties of different land cover types and the impact of urbanization on local climate characteristics. Accurate and continuous estimation across large spatial regions is crucial for the implementation of LST as an essential parameter in climate change mitigation strategies. Here, we propose a deep-learning-based methodology for LST estimation using multi-source data including Sentinel-2 imagery, land cover, and meteorological data. Our approach addresses common challenges in satellite-derived LST data, such as gaps caused by cloud cover, image border limitations, grid-pattern sensor artifacts, and temporal discontinuities due to infrequent sensor overpasses. We develop a regression-based convolutional neural network model, trained on ECOSTRESS (ECOsystem Spaceborne Thermal Radiometer Experiment on Space Station) mission data, which performs pixelwise LST predictions using 5 × 5 image patches, capturing contextual information around each pixel. This method not only preserves ECOSTRESS’s native resolution but also fills data gaps and enhances spatial and temporal coverage. In non-gap areas validated against ground truth ECOSTRESS data, the model achieves LST predictions with at least 80% of all pixel errors falling within a ±3 °C range. Unlike traditional satellite-based techniques, our model leverages high-temporal-resolution meteorological data to capture diurnal variations, allowing for more robust LST predictions across different regions and time periods. The model’s performance demonstrates the potential for integrating LST into urban planning, climate resilience strategies, and near-real-time heat stress monitoring, providing a valuable resource to assess and visualize the impact of urban development and land use and land cover changes. |
|---|---|
| ISSN: | 2072-4292 |