Deep transfer learning-based estimation of crop fractional vegetation cover at high spatiotemporal resolution using moderate-resolution satellite products

Fractional vegetation cover (FVC) is an essential parameter for assessing crop growth status and supporting precision agricultural management. Moderate-resolution FVC products support large-scale monitoring but lack field-scale spatial detail. High-resolution satellite imagery offers the potential to generate fine-resolution FVC estimates; however, its temporal continuity is often compromised by cloud contamination and revisit limitations. Furthermore, conventional FVC estimation approaches that rely solely on ground-truth data or radiative transfer simulations suffer from issues such as limited representativeness of field samples and the ill-posed nature of model inversion. To address these issues, this study proposes a data-fusion-based framework that integrates moderate-resolution FVC products, Sentinel-2 reflectance imagery, and deep transfer learning models. Specifically, temporally interpolated Global Land Surface Satellite (GLASS) and Geoland2 Version 3 (GEOV3) FVC products were matched with Sentinel-2 surface reflectance to generate robust training datasets. Long short-term memory (LSTM) and bidirectional LSTM (Bi-LSTM) models were employed to capture the complex relationship between FVC and reflectances. To mitigate potential biases originating from the source products, transfer learning techniques were applied to fine-tune the pretrained models using limited ground-measured FVC samples. The fine-tuned models were then used to generate high-spatiotemporal-resolution FVC maps based on reconstructed Sentinel-2 time-series imagery. The proposed method was benchmarked against a PROSAIL-based inversion, two machine learning algorithms, and vegetation index (VI)-based linear regression models. Results show that the fine-tuned Bi-LSTM model trained on FVC product-derived datasets achieved the highest accuracy (R2 = 0.8799, RMSE = 0.1170), outperforming random forest, XGBoost, and the model trained on PROSAIL simulations and VI-based models. Moreover, the transfer learning approach demonstrated superior spatial transferability across regions compared to conventional models. These findings highlight the effectiveness of leveraging moderate resolution FVC products and deep transfer learning to enhance FVC estimation at high spatiotemporal resolution, providing a scalable and reliable solution for crop monitoring applications.