Can High-Resolution Satellite Multispectral Imagery Be Used to Phenotype Canopy Traits and Yield Potential in Field Conditions?

Abstract

HighlightsVegetation indices (NDVI, GNDVI, and SAVI) extracted from high-resolution satellite imagery were significantly associated with vegetation indices extracted from UAV imagery.High-resolution satellite data can be used to predict maize yield at breeding plot scale.Breeding plot sizes and the variability between maize genotypes may be associated with prediction accuracies.Abstract. The recent availability of high spatial and temporal resolution satellite imagery has widened its applications in agriculture. Plant breeding and genetics programs are currently adopting unmanned aerial vehicle (UAV) based imagery data as a complement to ground data collection. With breeding trials across multiple geographic locations, UAV imaging is not always convenient. Hence, we anticipate that, similar to UAV imaging, phenotyping of individual test plots from high-resolution satellite imagery may also provide value to plant genetics and breeding programs. In this study, high spatial resolution satellite imagery (~38 to 48 cm pixel-1) was compared to imagery acquired using a UAV for its ability to phenotype maize grown in two-row and six-row breeding plots. Statistics (mean, median, sum) of color (red, green, blue), near-infrared, and vegetation indices such as normalized difference vegetation index (NDVI), green normalized difference vegetation index (GNDVI), and soil adjusted vegetation index (SAVI) were extracted from imagery from both sources (UAV and satellite) for comparison at three time points. In general, a strong correlation between satellite and UAV imagery extracted NDVI, GNDVI, and SAVI features (especially with mean and median statistics, p < 0.001) was observed at different time points. The correlation of both UAV and satellite image features with yield potential was maximum (p < 0.001) at the third time point (milk/dough growth stages). For example, Pearson’s correlation coefficients between mean NDVI, GNDVI, and SAVI features with yield potential were 0.52, 0.54, and 0.51 for data derived from UAV imagery, and 0.34, 0.41, and 0.40 for data derived from satellite imagery, respectively. Machine learning algorithms, including least absolute shrinkage and selection operator (Lasso) regression, were evaluated for yield prediction using vegetation index features that were significantly correlated with observed yield. The relationship between satellite imagery with crop performance can be a function of plot size in addition to crop variability. Nevertheless, with the ongoing improvement of satellite technologies, there is a possibility for the integration of satellite data into breeding programs, thus improving phenotyping efficiencies. Keywords: Image processing, Machine learning, Plant breeding, Statistical analysis, Unmanned aerial vehicles.