Skip to content

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

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

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

Authors: Solomiia Kurchaba, Angela Meyer Date: 2026-05-13 Paper ID: openalex:2605.13566

Summary

This paper addresses the spatiotemporal resolution trade-off in satellite-derived land surface temperature (LST) products by combining geostationary and polar-orbiting satellite data. A U-Net architecture is employed to downscale coarse-resolution data to 1 km at 15-minute intervals, which then serves as input for a ConvLSTM-based nowcasting model. The resulting framework provides accurate short-term intraday LST forecasts, outperforming standard statistical benchmarks in urban environments.

Key Contributions

  • Proposed a U-Net-based spatiotemporal downscaling framework that fuses SEVIRI and MODIS data to achieve 1 km spatial resolution at 15-minute intervals.
  • Developed a ConvLSTM-based nowcasting model for urban land surface temperatures with lead times of 15 to 75 minutes.
  • Achieved an RMSE of 1.92°C for downscaled LST estimation and outperformed persistence and climatological benchmarks for nowcasting (RMSE 0.57-1.15°C).

Open Questions & Future Work

Archivist Review

The paper provides a straightforward application of U-Net and ConvLSTM to LST downscaling and nowcasting. As these are standard architectures, no new concepts are approved. The open questions regarding seasonal generalization and physical driver integration were approved as they represent non-trivial research bottlenecks in remote sensing-based spatiotemporal forecasting.

Approved Open Questions

  • Year-round LST model generalization: Generalization across seasons is critical for the practical, continuous operational use of satellite-based LST monitoring for urban heat and climate applications.
  • Incorporating physical urban drivers: Explicit inclusion of physical drivers could improve model performance, robustness, and interpretability, especially in diverse urban environments where implicit encoding might fail.

Metadata & Links

url
https://arxiv.org/abs/2605.13566
paper_id
2605.13566
paper_source
openalex
domain
computer-vision
tags
architectures
datasets
concept_slugs
dataset_slugs
skill
TimeSeriesSkill
created_at
2026-05-16T06:02:18Z