Effects of Satellite Revisit Rate and Time-Series Smoothing Method on Throughout-Season Maize Yield Correlation Accuracy

Abstract

Predictions of crop yield made during the growing season aid in crop management and economic planning. Many yield prediction models are made by performing regression between satellite-derived vegetation indices (VI) and yield. This article studied the effects of time-series end date and satellite imaging frequency on the accuracy of VI-yield correlation. Daily, 3-m resolution, multispectral images were obtained over a maize field near Beltsville, MD, USA, in 2018 and 2019. Plot-average green normalized difference vegetation index (GNDVI) was extracted from these images. GNDVI time-series data were resampled to different revisit intervals, gap-filled and smoothed, temporally realigned, and correlated with plot-average yield at every day of the growing season. These experiments were then repeated with data removed from the end of the time-series. All methods tested performed well on time-series ending 72 d or more after green-up in 2019 (R-squared = 0.95) or time-series ending 65 d or more after green-up in 2018 (Flexfit R-squared = 0.92; shape model fitting R-squared = 0.89). All methods had poor correlation for time-series ending prior to the day of peak GNDVI. Mean R-squared values for GNDVI-yield correlations decreased with increasing revisit intervals. These trends were stronger in the 2019 data, with mean R-squared decreasing by more than 0.05 when sampled from 1 to 30-d revisit intervals (Flexfit) or to 22-d revisit intervals (shape model fitting). These findings, along with cloud-contamination statistics, were used to recommend an optimal methodology for yield correlation and an optimal overpass frequency of 1-4 d for future yield-monitoring satellite systems.