Sun et al. (2025) Dynamic monitoring of maize field vegetation cover using sentinel-1 and sentinel-2 data and transfer learning algorithms

Identification

• Journal: Smart Agricultural Technology
• Year: 2025
• Date: 2025-12-12
• Authors: Hongbo Sun, W. Liu, Miao Wang, Jinjin Li, R. Wang
• DOI: 10.1016/j.atech.2025.101711

Research Groups

• College of Resources and Environment, Shandong Agricultural University, Tai’an, China

Short Summary

This study developed a transfer learning model integrating multi-temporal Sentinel-1 and Sentinel-2 data with a pixel dichotomy model and temporal variation features to dynamically monitor maize fractional vegetation cover (FVC) during cloudy and rainy periods. The model demonstrated superior performance compared to classical machine learning methods, providing a robust solution for continuous crop monitoring under optical data limitations.

Objective

• To address the challenge of missing optical imagery during critical maize growth stages due to cloudy and rainy weather and the insufficient pre-training samples for transfer learning.
• To construct a mapping model between Sentinel-1 backscatter coefficients and Sentinel-2 FVC by incorporating temporal variation features of FVC across different growth stages into a transfer learning framework.
• To enable accurate reconstruction and prediction of summer maize FVC using only Sentinel-1 data, thereby providing technical support for continuous dynamic monitoring of summer maize FVC under complex meteorological conditions.

Study Configuration

• Spatial Scale: Local scale (Agricultural Science and Technology Innovation Park of Shandong Agricultural University, China) for training and initial validation, extended to a regional scale for large-area transferability validation.
• Temporal Scale: Maize growing season (approximately 90-120 days), covering seedling, jointing, trumpet, tasseling, flowering, silking, and maturity stages. Multi-year data (2019-2023) were used for Sentinel-1 and Sentinel-2 imagery.

Methodology and Data

• Models used:
  • Pixel dichotomy model (for FVC calculation from Sentinel-2 and UAV data).
  • Transfer learning algorithms based on a Conditional Generative Adversarial Network (CGAN) framework.
  • Generator: U-Net architecture with skip connections.
  • Discriminator: PatchGAN (Markovian discriminator) architecture.
  • L1 loss function for pixel-level similarity constraint.
  • Temporal variation features of FVC incorporated as prior constraints.
  • Random Forest (RF) regression model (for comparison).
  • Gaussian function (for fitting FVC time series).
• Data sources:
  • Satellite: Sentinel-1A SLC products (VV and VH polarizations, 6-day revisit cycle), Sentinel-2 Level-2A products (13 spectral bands, 10 m, 20 m, 60 m spatial resolutions, 5-day maximum revisit time).
  • UAV: DJI M600 Pro hexacopter with Sequoia multispectral sensor (green, red, red edge, near-infrared bands), 0.05 m spatial resolution.
  • DEM: SRTM DEM (for Sentinel-1 terrain correction).
  • Reference: Tianditu imagery (for large-area validation).

Main Results

• The proposed transfer learning model effectively captured dynamic FVC variations for both spring and summer maize across different phenological stages, demonstrating strong spatiotemporal consistency.
• Predicted FVC images showed strong consistency with real Sentinel-2 imagery in spatial patterns and vegetation growth dynamics, achieving an overall correlation of 0.75.
• The highest prediction accuracy was observed from the jointing to tasseling stage (R² = 0.55), with lower accuracies for the jointing–booting (R² = 0.35) and booting–tasseling (R² = 0.44) stages. Periods with larger FVC fluctuations corresponded to higher prediction accuracy.
• Compared to UAV-derived FVC, the model showed a tendency to overestimate FVC during the jointing stage, mixed deviations during booting, and systematic underestimation but improved performance (R² = 0.274) during the tasseling stage.
• For dynamic monitoring, the model achieved R² values of 0.4351 (jointing to tasseling), 0.5473 (jointing to silking), and 0.3453 (tasseling to silking), exhibiting a "shrinkage effect" where predictions were higher in low-difference areas and lower in high-difference areas.
• Large-area validation demonstrated strong generalizability, with predicted FVC spatial patterns closely aligning with regional land-use characteristics (e.g., road networks, rivers, bare land). A correlation coefficient of 0.88 was found between a predicted FVC map and a near-date Sentinel-2 image.
• The transfer learning framework significantly outperformed the Random Forest model, which showed very limited predictive performance (R² values from 0.000 to 0.013).
• The reconstructed FVC time series exhibited a typical "low–high–low" unimodal phenological curve, consistent with maize growth patterns, reaching a peak FVC close to 0.9 in mid-August.

Contributions

• Developed a novel approach for dynamic maize FVC monitoring by integrating microwave (Sentinel-1) and optical (Sentinel-2) remote sensing data with a transfer learning algorithm.
• Addressed the critical issue of optical data gaps during cloudy and rainy weather, enabling continuous FVC monitoring throughout the maize growing season.
• Enhanced the transfer learning framework by incorporating temporal variation features of FVC as prior knowledge, significantly reducing the model's reliance on large-scale labeled samples.
• Demonstrated the superior performance and generalizability of the proposed model compared to traditional machine learning methods (Random Forest) for FVC prediction.
• Provided insights into how different maize growth stages influence prediction performance, highlighting the model's sensitivity to canopy structural changes during rapid growth.
• Offered a practical, efficient, stable, and generalizable solution for AI-driven agricultural remote sensing applications, particularly for crop dynamic monitoring in regions prone to optical data limitations.

Funding

• Funds of the Natural Science Foundation of Shandong Province (ZR2020MD003)
• Shandong “Double Tops” Program (SYL2017XTTD02)
• Shandong Key R & D Program (2017CXGC0306)
• The Cultivate Plan Funds for Young Teacher and the Science and Technology Innovation Foundation for Youth of Shandong Agricultural University (23694)
• Shandong Provincial First-Class Discipline Construction Program

Citation

@article{Sun2025Dynamic,
  author = {Sun, Hongbo and Liu, W. and Wang, Miao and Li, Jinjin and Wang, R.},
  title = {Dynamic monitoring of maize field vegetation cover using sentinel-1 and sentinel-2 data and transfer learning algorithms},
  journal = {Smart Agricultural Technology},
  year = {2025},
  doi = {10.1016/j.atech.2025.101711},
  url = {https://doi.org/10.1016/j.atech.2025.101711}
}