The study, published in the Journal of Remote Sensing, integrates daily synthetic Harmonized Landsat Sentinel-2 (HLS) imagery with machine-learning models to reconstruct vegetation dynamics across agricultural fields. By analyzing vegetation index time series, the framework infers early crop development stages that are typically difficult to detect directly from space.
Understanding crop phenology - the timing of key developmental stages such as germination, growth, and senescence - is fundamental to agricultural management. Accurate crop calendars help optimize irrigation, fertilization, disease monitoring, and yield prediction. Traditional approaches depend on field observations or ground-based monitoring, but these are limited in spatial coverage and labor-intensive. Satellite remote sensing offers large-scale capabilities, yet detecting early stages like sowing and emergence is complicated by mixed soil-vegetation signals within satellite pixels, cloud cover, and data gaps.
The new framework addresses these challenges through an operational pipeline that combines satellite time-series reconstruction with phenological modeling. Continuous vegetation index data are first reconstructed from Landsat and Sentinel-2 imagery by filling gaps caused by cloud cover. From the reconstructed time series, six phenological stages - greenup, mid-greenup, maturity, senescence, mid-greendown, and dormancy - are extracted using an asymmetric double-sigmoid model. Machine-learning algorithms then infer sowing and emergence dates from the relationships between these stages.
Among the models tested, elastic net regression achieved the best performance, predicting sowing and emergence dates with an average error of approximately plus or minus 10 days. The approach was validated against ground observations from 20 PhenoCam monitoring sites across 13 U.S. states, producing a coefficient of determination (R2) of 0.94 and a bias of approximately 12 days.
The HLS dataset combines observations from Landsat 8/9 and Sentinel-2 satellites at a 30-meter spatial resolution. To handle cloud contamination, the team tested four gap-filling approaches - median interpolation, polynomial regression, harmonic modeling, and Light Gradient Boosting Machine (LightGBM). The polynomial method produced the most accurate reconstructions, preserving seasonal vegetation dynamics while minimizing noise.
"Our framework demonstrates that early crop development stages can be inferred indirectly from later phenological signals," the researchers noted. "Even though sowing and emergence are difficult to observe directly from satellite imagery, the seasonal growth trajectory contains enough information to reconstruct these dates."
The method was successfully applied across thousands of corn and soybean agricultural fields in the United States, demonstrating strong agreement with ground observations. The researchers note that with further refinement, the system could be integrated into global agricultural monitoring platforms and precision agriculture systems, contributing to food security and improved agricultural sustainability.
Research Report:Operational Framework for Field-Scale Crop Sowing and Emergence Date Estimation Using Daily Synthetic Harmonized Landsat Sentinel-2 Time Series
Related Links
Mississippi State University
Farming Today - Suppliers and Technology
| Subscribe Free To Our Daily Newsletters |
| Subscribe Free To Our Daily Newsletters |
