Liu et al. (2025) Estimating Soil Moisture Using Multimodal Remote Sensing and Transfer Optimization Techniques

Identification

• Journal: Remote Sensing
• Year: 2025
• Date: 2025-12-26
• Authors: Jingke Liu, Lin Liu, Weidong Yu, Xingbin Wang
• DOI: 10.3390/rs18010084

Research Groups

• Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing, China.
• School of Electronic, Electrical and Communication Engineering, University of Chinese Academy of Sciences, Beijing, China.
• School of Electronics and Information, Aerospace Information Technology University, Jinan, China.

Short Summary

This study develops a multimodal deep learning framework using a ConvNeXt v2 backbone and an intermediate fine-tuning strategy to estimate high-resolution surface soil moisture. By integrating SAR, optical, topographic, and meteorological data, the model achieved high precision ($R^2 = 0.8956$) and demonstrated robust transferability across diverse agro-ecological zones.

Objective

• To improve the accuracy, spatial resolution, and regional transferability of surface soil moisture (SSM) estimation by leveraging multimodal data fusion and a two-stage intermediate task fine-tuning strategy.

Study Configuration

• Spatial Scale: Field-scale (10 m resolution) across diverse regions in China, including Wangkui County (farmland), the Tibetan Plateau (alpine grassland), and the Heihe River Basin (environmentally fragile areas).
• Temporal Scale: Multi-year datasets including 2017–2020 for large-scale pre-training and various periods between 2010 and 2021 for in-situ validation.

Methodology and Data

• Models used: ConvNeXt v2 (backbone architecture) modified for regression, utilizing submanifold sparse convolutions, Squeeze-and-Excitation (SE) attention modules for multimodal fusion, and U-Net-style skip connections.
• Data sources:
  • SAR: Sentinel-1 (C-band) and Gaofen (L-band).
  • Optical: Sentinel-2 and Landsat-8 multispectral imagery.
  • Topography: ASTER GDEM (elevation and slope).
  • Contextual/Meteorological: MODIS (LAI/FPAR), ERA5 reanalysis (temperature and precipitation), and cyclic-encoded spatio-temporal metadata.
  • Labels: SMAP-WoSIS Fusion (SWF) for intermediate training and in-situ measurements (0–5 cm depth) for final fine-tuning.

Main Results

• High Accuracy: The multimodal framework achieved an $R^2$ of 0.8956, an RMSE of 4.5834%, and an MAE of 3.8412% when using the full suite of input variables.
• Fusion Benefits: Integrating optical and topographic data with SAR improved $R^2$ by approximately 0.20; adding meteorological and temporal context further increased $R^2$ by 0.05.
• Strategy Effectiveness: Intermediate fine-tuning on a large-scale dataset (10,571 images) followed by in-situ refinement reduced unbiased RMSE (ubRMSE) by 36.4% to 60.8% compared to non-fine-tuned models.
• Regional Robustness: In complex hydrological regions like the Heihe Basin and Tibet, multimodal fusion nearly halved the Mean Absolute Error (MAE) compared to SAR-only models (e.g., 0.1004 to 0.0509 in Heihe).

Contributions

• Intermediate Task Fine-Tuning: Proposed a two-stage training paradigm that bridges the gap between coarse-resolution satellite products and high-fidelity point measurements.
• Multimodal Architecture: Developed a fusion mechanism capable of processing heterogeneous pixel-level (raster) and image-level (scalar/metadata) inputs simultaneously.
• Field-Scale Monitoring: Demonstrated a scalable solution for generating 10 m resolution soil moisture maps, essential for precision irrigation and site-specific water management in heterogeneous landscapes.

Funding

• Aerospace Information Research Institute, Chinese Academy of Sciences (Grant number: Y9G0100BF0).

Citation

@article{Liu2025Estimating,
  author = {Liu, Jingke and Liu, Lin and Yu, Weidong and Wang, Xingbin},
  title = {Estimating Soil Moisture Using Multimodal Remote Sensing and Transfer Optimization Techniques},
  journal = {Remote Sensing},
  year = {2025},
  doi = {10.3390/rs18010084},
  url = {https://doi.org/10.3390/rs18010084}
}

Generated by BiblioAssistant using gemini-3-flash-preview (Google API)