Near real-time detection and forecasting of within-field phenology of winter wheat and corn using Sentinel-2 time-series data

Near real-time (NRT) crop phenology detection and forecasting at the sub-field level are important for crop growth monitoring and management in precision agriculture. Previous studies focused mainly on extracting phenological metrics (e.g., the start of season, and end of season) from time-series remote sensing data for a complete growing season. The existing NRT crop phenology detection methods are difficult to implement at the sub-field scale using available high spatial resolution satellite datasets with a low temporal resolution. In addition, these existing approaches can only estimate specific phenological events from the remote sensing perspective, which is different from the commonly used Biologische Bundesanstalt, Bundessortenamt and CHemical Industry (BBCH) scale used in crop phenology. In this study, an NRT phenology framework was proposed to detect and forecast phenology for winter wheat and corn from Sentinel-2 time-series data in Southwestern Ontario, Canada. The framework incorporated both the canopy structure dynamics model (CSDM) and the shape model-fitting approach to capture crop growth of development over time. The framework can be performed in NRT using timely available Sentinel-2 data during the growing season. The day of year (DOY) for each phenological stage and the BBCH scale on a specific date can be obtained. The resultant Root Mean Squared Errors (RMSEs, days) of BBCH scales were less than 2.9 for winter wheat and less than 3.7 for corn. The RMSEs of detected DOY of all the phenological stages were less than 4 days for winter wheat and less than 3.7 days for corn except for the senescence and the end of season (EOS) stage.

Automated summary

Near real-time (NRT) crop phenology detection and forecasting is crucial for precision agriculture. This study proposes a framework using Sentinel-2 time-series data to detect and forecast phenology for winter wheat and corn. The framework incorporates both canopy structure dynamics model (CSDM) and shape model-fitting approach. The results show that the framework can achieve accurate detection and forecasting with low RMSEs.