Limitations and solutions for developing a grain yield and protein content forecasting model based on vegetation indices in organic wheat production – on-farm experimentation

Abstract

ABSTRACT Multispectral imaging has gained increased attention in recent years for monitoring and predicting crop-related characteristics. Vegetation indices can provide valuable insight into crop health and growth, especially in conventional agriculture and small-sized experimental plots. However, the effectiveness of such methods in non-conventional agricultural systems still needs to be clarified. This study was the first to test three vegetation indices (NDVI, GNDVI, NDRE) for grain yield and protein content prediction in organic farming under real-life farm management practices. Using two small-plot and one on-farm organic wheat variety trial sites, unmanned aerial vehicle-based remote sensing data was compared with physical sampling results in 2 years to assess the potential of agricultural remote sensing to predict yield and protein content on small- and medium-sized plots of organic winter wheat (Triticum aestivum L.). The correlations between vegetation indices and crop traits were stronger in small- than in medium-sized plots (r max = 0.85 and 0.75, respectively, for grain yield), probably due to increased heterogeneity of the larger plots, due to differences in weed cover, soil nutrient supply, etc., and it was strongly affected by the time of measurement. Linear equation for yield and protein content prediction was not sufficiently precise; the machine learning algorithms improved the accuracy of yield estimation. Both methods considered NDRE with a higher weight than NDVI when modelling grain yield and protein content based on the vegetation index values, which suggested that NDRE was more suitable for prediction. Small-plot-based predictions over-estimated grain yield and under-estimated grain protein content in medium-sized plots.