Heidarian et al. (2025) Cross-Domain Land Surface Temperature Retrieval via Strategic Fine-Tuning-Based Transfer Learning: Application to GF5-02 VIMI Imagery

Identification

• Journal: Remote Sensing
• Year: 2025
• Date: 2025-11-23
• Authors: Peyman Heidarian, Hua Li, Zelin Zhang, Yumin Tan, Feng Zhao, Biao Cao, Yongming Du, Qinhuo Liu
• DOI: 10.3390/rs17233803

Research Groups

• Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing 100101, China
• Hangzhou International Innovation Institute, Hangzhou 311115, China
• School of Transportation Science and Engineering, Beihang University, Beijing 100191, China
• School of Instrumentation Science and Opto-Electronics Engineering, Beihang University, Beijing 100191, China
• State Key Laboratory of Earth Surface Processes and Disaster Risk Reduction, Advanced Interdisciplinary Institute of Satellite Applications, Faculty of Geographical Science, Beijing Normal University, Beijing 100875, China

Short Summary

This study introduces a three-stage strategic fine-tuning-based transfer learning (SFTL) framework to retrieve Land Surface Temperature (LST) from GF5-02 VIMI imagery, demonstrating superior cross-site generalization (RMSE ≈ 2.89–3.34 K) compared to traditional methods by integrating large simulated datasets with limited in situ measurements and parameter-efficient fine-tuning strategies.

Objective

• To develop and validate a three-stage strategic fine-tuning-based transfer learning (SFTL) framework for accurate and generalizable Land Surface Temperature (LST) retrieval from GF5-02 Visible and Infrared Multispectral Imager (VIMI) imagery, addressing data scarcity, regional variability, and limited input variables.

Study Configuration

• Spatial Scale: Heihe and Huailai regions in China; GF5-02 VIMI imagery with 40 m spatial resolution for thermal infrared (TIR) channels. Ground sites cover diverse landforms including cropland, wetland, desert steppe, shrub, forest, and crop.
• Temporal Scale: In situ measurements from 2021–2023. GF5-02 VIMI revisit period up to 51 days.

Methodology and Data

• Models used:
  • Strategic Fine-Tuning-based Transfer Learning (SFTL) framework incorporating:
    • Deep Neural Network (DNN)
    • Transformer (TrF)
    • Convolutional Neural Network (CNN)
    • Random Forest (RF)
    • Light Gradient Boosting Machine (LGBM)
  • Split-Window (SW) algorithm (physics-based baseline)
  • Fine-tuning strategies: Full, Head, Gradual, Adapter, Low-Rank Adaptation (LoRA).
• Data sources:
  • Satellite: GF5-02 Visible and Infrared Multispectral Imager (VIMI) Level-1 imagery (TIR channels centered at 8.20, 8.63, 10.80, and 11.95 µm, 40 m resolution).
  • Observation (in situ): 18 ground measurement sites (3 in Heihe, 15 in Huailai, China) equipped with net radiometers (CNR1, Kipp and Zonen CGR3) for LST validation (2021–2023).
  • Simulated: Thermodynamic Initial Guess Retrieval 2000 (TIGR 2000) database (946 clear-sky atmospheric profiles) combined with ASTER spectral library (81 emissivity spectra) and MODTRAN 5.2 radiative transfer model, generating 430,596 samples.
  • Reanalysis: ERA5 reanalysis data for Water Vapor Content (WVC) (~25 km resolution).
  • Auxiliary: ASTER Global Emissivity Dataset (GED) v3 for Land Surface Emissivity (LSE) estimation using the Vegetation Cover Method (VCM).

Main Results

• The SFTL framework achieved strong cross-site generalization, outperforming the Split-Window (SW) algorithm and direct-training machine learning models.
• Huailai-to-Heihe transfer (SiHuHe): DNN with gradual fine-tuning achieved the best performance with an in-domain RMSE of ≈2.62 K and cross-site RMSE of ≈2.89 K (R² ≈ 0.96) on Heihe.
• Heihe-to-Huailai transfer (SiHeHu): Transformer (TrF) with head-only fine-tuning achieved an in-domain RMSE of ≈2.56 K and cross-site RMSE of ≈3.34 K (R² ≈ 0.94) on Huailai. Parameter-efficient methods like TrF-LoRA (cross-site RMSE ≈ 3.41 K) and CNN-LoRA (cross-site RMSE ≈ 3.67 K) also performed well.
• The traditional Split-Window (SW) algorithm yielded RMSEs of 3.67 K for Huailai and 4.22 K for Heihe. Direct-training ML models showed substantial performance degradation under domain shift (RMSEs > 4.17 K cross-site).
• Application to real GF5-02 VIMI imagery confirmed the best SFTL configurations, achieving RMSEs around 2.24–3.5 K and R² ≈ 0.88–0.96, with minimal bias (absolute bias ≤ 3.03 K in Huailai, < 0.99 K in Heihe).
• The feature-engineered variable WVC × ΔBT was pivotal in pre-training, significantly improving accuracy (e.g., TrF RMSE increased by 6.845 K and R² reduced by 0.062 upon its removal).

Contributions

• Introduces a novel three-stage Strategic Fine-Tuning-based Transfer Learning (SFTL) framework for LST retrieval, combining large-scale radiative transfer simulations, an engineered humidity-sensitive feature (WVC × ΔBT), and multiple parameter-efficient fine-tuning strategies.
• Demonstrates superior cross-site generalization capabilities (RMSE ≈ 2.89–3.34 K) compared to the widely used Split-Window algorithm and direct-training machine learning models, particularly in data-scarce regions.
• Provides a scalable and data-efficient solution for operational GF5-02 VIMI LST retrieval across heterogeneous surface types and atmospheric conditions, reducing dependence on extensive labeled in situ datasets.
• Evaluates and identifies optimal neural network architectures and fine-tuning strategies (e.g., DNN-gradual for larger target domains, TrF-head/LoRA for smaller target domains) based on in situ sample size and statistical variability.
• Validates the framework using real GF5-02 VIMI imagery and spatio-temporally matched in situ observations, confirming its robustness and physical plausibility.

Funding

• National Key Research and Development Program of China (Grant 2023YFB3905801)
• National Natural Science Foundation of China (Grant 42471365)

Citation

@article{Heidarian2025CrossDomain,
  author = {Heidarian, Peyman and Li, Hua and Zhang, Zelin and Tan, Yumin and Zhao, Feng and Cao, Biao and Du, Yongming and Liu, Qinhuo},
  title = {Cross-Domain Land Surface Temperature Retrieval via Strategic Fine-Tuning-Based Transfer Learning: Application to GF5-02 VIMI Imagery},
  journal = {Remote Sensing},
  year = {2025},
  doi = {10.3390/rs17233803},
  url = {https://doi.org/10.3390/rs17233803}
}