High-resolution L-band VOD retrieval using GNSS-T–calibrated multi-sensor fusion and self-supervised learning

Vegetation Optical Depth (VOD) from L-band microwave radiometry is widely used to monitor vegetation water status and ecosystem dynamics, but existing SMAP and SMOS products at 25–40 km resolution cannot resolve fine-scale canopy heterogeneity. We develop and evaluate a framework for estimating forest Global Navigation Satellite System Transmissometry (GNSS-T)-calibrated VOD at 30 m resolution by fusing Sentinel-1A Synthetic Aperture Radar (SAR), Harmonized Landsat–Sentinel-2 optical reflectance, GEDI lidar canopy structure, and climate variables (precipitation and air temperature aggregated over same-day, 7-, 30-, and 60-day windows), all calibrated using ground-based (GNSS-T). A self-supervised learning (SSL) framework based on Vision Transformers, pretrained on unlabeled satellite imagery and fine-tuned with GNSS-derived VOD from five forest sites in the United States and Switzerland, generally outperforms Random Forest and XGBoost across SAR-only, optical-only, and fusion configurations. Feature importance analysis reveals that medium-term climate drivers, particularly 60-day precipitation and 7-day temperature, consistently outrank SAR backscatter, while SAR captures rain-driven VOD peaks that optical sensors miss due to cloud screening. The framework demonstrates robust temporal generalization across deciduous and evergreen forest types, though spatial generalization remains the critical next step as the GNSS-T network expands. These results establish GNSS-T-calibrated multi-sensor fusion with domain-specific SSL pretraining as a viable pathway toward operational 30 m L-band VOD estimation.