Progressive Early-Season Crop Mapping over the Contiguous United States Using Sentinel-2 Time Series and Historical Ground Data

Timely early-season crop maps can provide valuable spatial information on crop distribution. However, generating such maps from satellite imagery is challenging due to limited observations and inaccessible ground data for the current year. In this study, we evaluated the potential of using Sentinel-2 time series and historical ground data to map crops progressively from early to late season across the entire Contiguous United States (CONUS). We processed all available Sentinel-2 imagery during the maize and soybean growing season (May through October) from 2019 to 2022, producing 10-day composites as Sentinel-2 Analysis Ready Data (ARD). Ground data were collected over random sample locations across CONUS from 2019 to 2022. Focusing on 2022 as an experimental year, we implemented a progressive crop mapping method at 10-day intervals throughout the season. Random Forest classifiers were trained on prior years’ data (e.g., 2019–2021) and applied to the 2022 Sentinel-2 time series. We designed three training scenarios using different combinations of historical ground data to assess how their availability affects national-scale early- and in-season crop classification. For each interval and scenario, Random Forest models were separately trained for maize and soybean, and the resulting 10-day maps were validated using 2022 ground reference. Our results show that greater availability of historical ground data consistently improved classification accuracy. When using three years of training data (2019–2021), the 2022 early-season maps achieved F1 scores exceeding 80% for both maize and soybean by July 19. In addition, we compared each 10-day map to a reference end-of-season crop map (Li et al., 2025) to determine the earliest day-of-year (DOY) when a pixel’s classification matched the reference and remained stable for the rest of the season. This analysis revealed that 50% of maize and soybean pixels could be accurately identified by late June across over 50% of U.S. counties. Our findings demonstrate the potential of generating timely early-season crop maps using open satellite data and historical ground observations, providing critical information for policymakers, farmers, and private stakeholders in support of food supply chain monitoring, disaster response, and agricultural production forecasting.

Li, H., Song, X.P., Adusei, B., Potapov, P., Hansen, M. and Gao, F. (2025), Progressive Early-Season Crop Mapping over the Contiguous United States Using Sentinel-2 Time Series and Historical Ground Data. AGU Fall Meeting, Abstracts GC43R-1033, December 14-19, New Orleans, Louisiana, USA.