Forecasting Maize Yield from Different Spatial Resolutions of Vegetation Indices Using Sentinel-2 Imagery and R

Abstract

Vegetation indices derived from freely available satellite imagery have strong potential for crop yield modeling. However, positional inaccuracies and scale mismatches between yield monitor data and satellite imagery, together with noise at native spatial resolutions, can limit their practical applicability. Therefore, this study evaluated the predictive performance of multiple coarser spatial resolutions (10 m to 60 m), generated through pixel aggregation to attenuate spatial noise, for maize yield prediction. Sentinel-2 imagery acquired at phenological stages V8, R1, and R5 was used to compute NDVI, ARVI, and GNDVI vegetation indices. Additionally, two methods for generating raster yield maps were evaluated: mean-based rasterization and focal window aggregation considering harvester swath width. Predictive models were evaluated using spatial k-fold cross-validation. The GNDVI index showed the highest sensitivity, and the R1 stage provided the best predictive performance. Results demonstrated increased correlation and reduced prediction errors with coarser spatial resolutions, indicating effective noise attenuation through spatial aggregation. Mean yield prediction errors decreased from 0.93 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Mg~ha^{-1}$$\end{document} (9.7%) at 10 m resolution to 0.60 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Mg~ha^{-1}$$\end{document} (6.3%) at 40 m, while the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} increased from 0.24 to 0.52. At the 10 m native resolution, the focal method increased the prediction \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} by roughly 23% compared to the mean rasterization method. Spatial aggregation significantly enhanced data usability for practical applications. The resulting models demonstrated potential for large-scale yield estimation and management zone delineation, while also providing a foundation for the development of more advanced predictive modeling approaches using available Sentinel-2 imagery.