Mapping grain crop sowing date in smallholder systems using optical imagery

Abstract

Sowing date prediction using Earth observation data is challenging in smallholder systems due to small field sizes, heterogeneity in management practices, and a lack of reference data. This study aims to develop a generalizable algorithm that does not require any ground data for calibration to map sowing date using the Normalized Difference Vegetation Index (NDVI) from three optical datasets: MODIS, Harmonized Landsat and Sentinel (HLS), and Sentinel-2. We applied Savitzky-Golay (SG) and spline smoothing algorithms to each dataset and developed a derivative approach to identify the inflection point that represents the Start of Season (SoS), which was then converted to sowing date. We applied our methodology to map the sowing date of winter wheat in Bihar, India and spring-summer maize in the state of Mexico, Mexico. Overall, Sentinel-2 data led to the highest accuracies, but the performance of the smoothing algorithm differed across locations. In India, prediction models using SG achieved an R² of 0.45 and a root mean square deviation (RMSD) of 11.44 days. In Mexico, prediction models using spline performed best, with an R² of 0.19 and an RMSD of 4.24 weeks. The lower accuracy in Mexico was due to more complex cropping patterns as well as noise in the observed sowing date dataset. Our algorithm shows potential to identify SoS, and ultimately sowing date, at scale using Sentinel-2 imagery. However, challenges from low-quality validation datasets, small field sizes, cloud cover, and landscape complexity continue to pose challenges to predict sowing date using Earth observation data products.