Shen et al. (2026) 30-meter land surface temperature from Landsat via progressive self-training downscaling

Identification

• Journal: Remote Sensing of Environment
• Year: 2026
• Date: 2026-10-21
• Authors: Huanfeng Shen, Chan Li, Menghui Jiang, Penghai Wu, Guanhao Zhang, Tian Xie
• DOI: 10.1016/j.rse.2026.115732

Research Groups

• School of Resource and Environmental Sciences, Wuhan University, Wuhan, 430079, China
• Key Laboratory of Geographic Information System of Ministry of Education, Wuhan, 430079, China
• Key Laboratory of Digital Cartography and Land Information Application of the Ministry of Natural Resources, Wuhan, 430079, China
• Anhui Province Key Laboratory of Wetland Ecosystem Protection and Restoration, Anhui University, Hefei, Anhui, 230601, China

Short Summary

This study introduces a progressive self-training framework with a cross-modal fusion network to downscale Landsat Land Surface Temperature (LST) from 100 m to 30 m resolution, achieving enhanced accuracy and spatial fidelity without requiring fine-scale ground truth. The framework significantly improves LST reconstruction compared to existing methods, providing valuable data for fine-scale environmental studies.

Objective

• To develop a progressive self-training framework for downscaling Landsat LST to 30 m spatial resolution, without relying on fine-scale ground truth and with minimal external data dependence, to meet the demands of fine-scale surface process studies and localized environmental monitoring.

Study Configuration

• Spatial Scale: Downscaling from 100 m (native thermal band resolution) to 30 m (target resolution, matching optical bands).
• Temporal Scale: Designed for Landsat series, which provides over 40 years of consistent LST records.

Methodology and Data

• Models used: Progressive self-training framework, cross-modal fusion network, cross-modal dictionary attention mechanism, hybrid CNN-Transformer architecture, frequency-domain guidance strategy (implemented at network structure and loss-function levels), high-frequency structural loss.
• Data sources: Landsat LST (100 m native resolution), Landsat optical bands (30 m, as auxiliary variables), SDGSAT-1 30 m LST (for spatial validation), in situ measurements (for temporal validation).

Main Results

• The proposed framework effectively reconstructs fine-scale thermal patterns while preserving spatial heterogeneity.
• Ablation analyses confirmed that the fusion design and auxiliary variable combination effectively reduce downscaling errors.
• The frequency-domain guidance strategy, particularly the high-frequency structural loss, was effective in reconstructing crisp fine-scale thermal textures.
• Spatial validation against SDGSAT-1 30 m LST and temporal validation using in situ measurements yielded a station-averaged Root Mean Square Error (RMSE) of 3.465 K and a coefficient of determination (R²) of 0.909.
• This performance significantly surpasses the official cubic interpolation product, which showed an RMSE of 4.195 K and an R² of 0.861.

Contributions

• Proposes a novel progressive self-training framework for downscaling Landsat LST to 30 m without the need for fine-scale ground truth.
• Introduces a cross-modal fusion network that leverages a cross-modal dictionary attention mechanism and a hybrid CNN-Transformer architecture to effectively extract complementary information between LST and auxiliary variables.
• Incorporates a frequency-domain guidance strategy to enhance structural fidelity and suppress artifact accumulation during progressive refinement.
• Provides a stable and accurate pathway for enhancing the spatial resolution of Landsat LST, offering improved data support for fine-scale surface process studies and localized environmental monitoring.

Funding

Not explicitly stated in the provided text.

Citation

@article{Shen202630meter,
  author = {Shen, Huanfeng and Li, Chan and Jiang, Menghui and Wu, Penghai and Zhang, Guanhao and Xie, Tian},
  title = {30-meter land surface temperature from Landsat via progressive self-training downscaling},
  journal = {Remote Sensing of Environment},
  year = {2026},
  doi = {10.1016/j.rse.2026.115732},
  url = {https://doi.org/10.1016/j.rse.2026.115732}
}