In-season crop progress in unsurveyed regions using networks trained on synthetic data
Worrall and Judge (2026)
How can we monitor crop development in regions where no ground observations exist?
Many remote sensing approaches for crop progress rely on survey data, limiting their use in data-sparse regions. As a result, these approaches have limited transferability across regions with different climates and management practices.
Worrall and Judge (2026) address this problem by developing a framework that combines satellite observations with synthetic crop progress data to enable in-season monitoring in unsurveyed regions. The study links weather generators, crop growth models (DSSAT), and radiative transfer models to simulate realistic crop development and associated reflectance signals. A bi-directional LSTM neural network is then trained using combinations of real survey data from the U.S. Midwest and synthetic data generated for Argentina, which serves as a proxy “unsurveyed” region.
Their findings show that:
- Including synthetic data during training improves crop progress estimation in unsurveyed regions, increasing overall performance by ~8.7% compared to using surveyed data alone.
- Performance gains are strongest in regions with more complex cropping systems, such as dual planting seasons, where synthetic data helps the model capture bimodal crop development patterns.
- Combining synthetic and real data also reduces sensitivity to noise in satellite NDVI time series, improving robustness of in-season estimates.