Zhang et al. (2025) A seamless global daily 5 km soil moisture product from 1982 to 2021 using AVHRR satellite data and an attention-based deep learning model

Identification

• Journal: Earth system science data
• Year: 2025
• Date: 2025-10-07
• Authors: Yufang Zhang, Shunlin Liang, Han Ma, Tao He, Feng Tian, Guodong Zhang, Jianglei Xu
• DOI: 10.5194/essd-17-5181-2025

Research Groups

• School of Software, Northwestern Polytechnical University, Xi’an, China
• Department of Geography, University of Hong Kong, Hong Kong SAR, China
• Hubei Key Laboratory of Quantitative Remote Sensing of Land and Atmosphere, School of Remote Sensing and Information Engineering, Wuhan University, Wuhan, China
• Faculty of Geosciences and Environmental Engineering, Southwest Jiajiaotong University, Chengdu, China

Short Summary

This study generated a consistent and seamless global daily 5 km surface soil moisture (0–5 cm) product spanning 1982–2021 using an attention-based deep learning model (AtLSTM) trained with AVHRR satellite data and other multi-source inputs. The resulting GLASS-AVHRR SM product demonstrates superior accuracy, spatiotemporal completeness, and richer spatial details compared to existing long-term global soil moisture datasets.

Objective

• To develop a deep learning (DL)-based global soil moisture (SM) estimation model by integrating multi-source datasets and leveraging their complementary strengths to derive a seamless and reliable long-term global SM product.
• To compare the performance of different DL models (basic LSTM, Bi-LSTM, AtLSTM, and transformer) with the benchmark eXtreme Gradient Boosting (XGBoost) model and to investigate the effect of input sequence length on model accuracy.
• To fully evaluate the accuracy and spatiotemporal consistency of the derived long-term GLASS-AVHRR SM product through validation against in situ SM datasets across different spatial scales and intercomparison with other long-term global SM products.

Study Configuration

• Spatial Scale: Global, 5 km spatial resolution, representing the uppermost soil layer (0–5 cm).
• Temporal Scale: Daily, spanning 40 years from 1982 to 2021.

Methodology and Data

• Models used:
  • Deep Learning Models: Long Short-Term Memory (LSTM), Bidirectional LSTM (Bi-LSTM), Attention-based LSTM (AtLSTM), Transformer.
  • Benchmark Model: eXtreme Gradient Boosting (XGBoost).
• Data sources:
  • Input Features:
    • GLASS-AVHRR albedo (black-sky visible, near-infrared, shortwave) and Land Surface Temperature (LST) products (5 km).
    • ERA5-Land reanalysis SM product (0.1°, 0–7 cm).
    • MERIT Digital Elevation Model (DEM) for elevation, slope, and aspect (90 m, resampled to 5 km).
    • SoilGrids datasets for sand, silt, and clay content (250 m, 0–5 cm, resampled to 5 km).
  • Training Target: GLASS-MODIS SM product (1 km, 0–5 cm, 2000–2020, resampled to 5 km).
  • Validation Data:
    • International Soil Moisture Network (ISMN) in situ SM (point-scale, 0–5 cm, 1982–1999 and 2000–2018).
    • COsmic-ray Soil Moisture Observing System (COSMOS) networks (field-scale, 130–240 m footprint, 15–83 cm depth, post-2000).
    • Soil Moisture Active Passive (SMAP) core validation sites (CVS) in situ SM (upscaled 9 km, 0–5 cm, 2015–2021).
  • Intercomparison Data:
    • European Space Agency Climate Change Initiative (ESA CCI) combined SM product (0.25°, 1978–2021).
    • ERA5-Land SM product (0.1°, 1950–present).

Main Results

• All four deep learning models (LSTM, Bi-LSTM, AtLSTM, Transformer) outperformed the benchmark XGBoost model on the test set, especially at high soil moisture (SM) levels (> 0.4 m³ m⁻³).
• The Attention-based LSTM (AtLSTM) model achieved the highest accuracy with a coefficient of determination (R²) of 0.987 and a root mean square error (RMSE) of 0.011 m³ m⁻³ on the test set. For high SM levels (> 0.4 m³ m⁻³), AtLSTM achieved R² of 0.621 and RMSE of 0.016 m³ m⁻³.
• The AtLSTM model effectively learned short-term adjacent temporal dependencies, with its overall accuracy stabilizing at an input sequence length of approximately 4 days.
• The generated GLASS-AVHRR SM product showed high accuracy when validated against independent in situ data:
  • ISMN stations (1982–1999): median R of 0.73 and median unbiased RMSE (ubRMSE) of 0.041 m³ m⁻³.
  • COSMOS networks (post-2000): median R values ranging from 0.63 to 0.79 and median ubRMSE values from 0.044 to 0.065 m³ m⁻³.
  • SMAP CVSs (2015–2021): overall R of 0.82 and ubRMSE of 0.054 m³ m⁻³.
• The GLASS-AVHRR SM product successfully corrected the large wet biases observed in the input ERA5-Land SM product.
• The product exhibited the most complete spatial coverage, contained much richer spatial details, and showed high spatial and temporal consistency with the GLASS-MODIS SM product, outperforming the coarser resolution ERA5-Land and ESA CCI SM products.

Contributions

• Generation of the first four-decade (1982–2021) seamless global daily surface soil moisture product (GLASS-AVHRR SM) at 5 km resolution, leveraging long-archived AVHRR satellite observations and deep learning.
• Demonstrated the superior performance of the attention-based LSTM (AtLSTM) model for soil moisture estimation, particularly its ability to capture critical short-term temporal dependencies and improve accuracy at high soil moisture levels.
• Provided a valuable extension to the GLASS-MODIS SM product and a robust complement to microwave-based SM products, offering complete spatial coverage, reliable accuracy, and consistency over a long period, addressing limitations of existing datasets (e.g., spatial gaps, low resolution, biases).

Funding

• Open Research Program of the International Research Center of Big Data for Sustainable Development Goals (grant no. CBAS2022ORP01)
• National Key Research and Development Program of China (grant no. 2023YFF1303702)
• Fundamental Research Funds for the Central Universities (grant no. G2025KY05116)
• National Key Research and Development Program of China (grant no. 2016YFA0600103)
• National Natural Science Foundation of China (grant no. 42090011)

Citation

@article{Zhang2025seamless,
  author = {Zhang, Yufang and Liang, Shunlin and Ma, Han and He, Tao and Tian, Feng and Zhang, Guodong and Xu, Jianglei},
  title = {A seamless global daily 5 km soil moisture product from 1982 to 2021 using AVHRR satellite data and an attention-based deep learning model},
  journal = {Earth system science data},
  year = {2025},
  doi = {10.5194/essd-17-5181-2025},
  url = {https://doi.org/10.5194/essd-17-5181-2025}
}