Qiao et al. (2026) High spatial resolution GLASS FAPAR (version 2) product from Landsat imagery: Algorithm development using a knowledge transfer strategy

Identification

• Journal: International Journal of Applied Earth Observation and Geoinformation
• Year: 2026
• Date: 2026-01-06
• Authors: Yuting Qiao, Huaan Jin, Tao He, Shunlin LIANG, Tian Feng, Wei Zhao, Zhouyang Liu
• DOI: 10.1016/j.jag.2025.105051

Research Groups

• Research Center of Digital Mountain and Remote Sensing Application, Institute of Mountain Hazards and Environment, Chinese Academy of Sciences, Chengdu, China
• University of Chinese Academy of Sciences, Beijing, China
• School of Remote Sensing and Information Engineering, Wuhan University, Wuhan, China
• Department of Geography, University of Hong Kong, Hong Kong, China

Short Summary

This study develops and validates the Hi-GLASS FAPAR (version 2) product, a high spatial resolution (30 m) fraction of absorbed photosynthetically active radiation (FAPAR) product derived from Landsat imagery, by integrating deep transfer learning with radiative transfer models to enhance accuracy and adaptability, especially over heterogeneous surfaces. The new product significantly outperforms its predecessor and other coarse-resolution products, offering improved spatial detail and temporal consistency.

Objective

• To introduce deep learning and transfer learning techniques to update the current Hi-GLASS FAPAR V1 algorithm and generate the corresponding version 2 product.
• To rigorously evaluate the Hi-GLASS FAPAR V2 accuracy and reliability through comparisons with field measurements and existing FAPAR products.

Study Configuration

• Spatial Scale: 30 m (Landsat imagery, Hi-GLASS FAPAR V2 product); 3 km × 3 km (DIRECT V2.1 reference maps); 250 m (GLASS FAPAR); 500 m (MODIS FAPAR); 1/112° (GEOV2 FAPAR).
• Temporal Scale: 16-day revisit interval (Landsat); 2000–2021 (DIRECT V2.1 database); 2015–2020 (time series analysis for selected sites).

Methodology and Data

• Models used:
  • Deep Learning: Long Short-Term Memory (LSTM) model.
  • Radiative Transfer Model (RTM): Soil-Leaf-Canopy (SLC) model (integrates GSV soil reflectance model, SAIL canopy reflectance model, PROSPECT-D leaf optical model).
  • Transfer Learning (TL) strategy.
• Data sources:
  • Satellite: Landsat surface reflectance data (TM, ETM+, OLI sensors) from USGS EarthExplorer.
  • Observation: Publicly available ground-truth FAPAR data from ImagineS, VALERI, and GBOV networks.
  • Reference/Validation: DIRECT V2.1 database (3 km × 3 km spatially averaged FAPAR values), GEOV2, GLASS, and MODIS FAPAR products.

Main Results

• The Hi-GLASS FAPAR V2 product achieved an overall accuracy of R² = 0.95 and RMSE = 0.08, significantly outperforming V1 (R² = 0.94, RMSE = 0.11) and the pre-trained LSTM FAPAR (R² = 0.95, RMSE = 0.09) against in-situ observations, with an Uncertainty Agreement Ratio (UAR) of 86%.
• The greatest improvement in FAPAR retrieval accuracy was observed over multiple forest types, with R² increasing by 2%–11% and RMSE decreasing by 15%–55%, demonstrating enhanced adaptability to heterogeneous canopies.
• Hi-GLASS V2 preserved better spatial details and exhibited temporal dynamics more closely aligned with field measurements compared to V1 and coarse-resolution products (MODIS, GLASS, GEOV2).
• When compared against DIRECT2.1 reference values, Hi-GLASS FAPAR V2 showed superior accuracy (R² = 0.84, RMSE = 0.10, bias = -0.01) relative to GLASS, MODIS, and GEOV2 FAPAR products.
• Transfer learning models demonstrated insensitivity to the size of the target domain training samples (10%–90%), maintaining stable generalization performance.
• The efficacy of the transfer learning model was linked to the distribution of training data, with models trained on crops or grasslands performing best, while those trained on forests showed systematic overestimation in shrub areas.

Contributions

• Developed an updated Hi-GLASS FAPAR V2 algorithm and product by integrating deep transfer learning and radiative transfer models, addressing limitations of the previous version.
• Achieved significant improvements in FAPAR retrieval accuracy and adaptability to heterogeneous surfaces, particularly for forest types, through a physical knowledge-guided transfer learning approach.
• Generated a high spatial resolution (30 m) FAPAR product with long temporal coverage from Landsat imagery, providing enhanced spatial details and temporal consistency compared to existing coarse-resolution products.
• Demonstrated the effectiveness of transfer learning in mitigating data scarcity in the target domain and improving model capacity for specific feature extraction across diverse vegetation types.
• Provided a robust and generalizable method for FAPAR estimation, offering higher-quality data support for applications such as ecosystem carbon cycling and vegetation growth monitoring.

Funding

• National Key Research and Development Program of China (Grant No. 2020YFA0608704)
• National Natural Science Foundation of China (Grant No. 42571455, 42222109)
• Sichuan Science and Technology Program (Grant No. 2024NSFSC0077)

Citation

@article{Qiao2026High,
  author = {Qiao, Yuting and Jin, Huaan and He, Tao and LIANG, Shunlin and Feng, Tian and Zhao, Wei and Liu, Zhouyang},
  title = {High spatial resolution GLASS FAPAR (version 2) product from Landsat imagery: Algorithm development using a knowledge transfer strategy},
  journal = {International Journal of Applied Earth Observation and Geoinformation},
  year = {2026},
  doi = {10.1016/j.jag.2025.105051},
  url = {https://doi.org/10.1016/j.jag.2025.105051}
}