Prediction modeling for yield and water-use efficiency in spinach using remote sensing via an unmanned aerial system

Abstract

Predictive modeling (PM) is a useful tool in selecting from a large number of target variables. Given the large quantities of data types generated from unmanned aerial system (UAS) platforms, uncovering the most appropriate candidate UAS data–based phenotype (UASDP) with a strong relationship to target crop traits may be a challenge with the traditional regression models. We hypothesized that by employing machine learning modeling techniques, the contributions of multiple vegetative indices as predictors of biomass and water-use could be defined and ranked. The objective of this study was to apply predictive machine techniques in determining the phase of spinach growth and UASDP variables that best predict yield and water-use efficiency. UASDPs were derived using red-green-blue and multi-spectral sensors mounted on a UAS platform flown weekly on 10 spinach genotypes under well-watered and partial water deficit conditions. Candidate UASDPs including mean and maximum plant height, canopy cover and volume, excess greenness index (ExG), chlorophyll red-edge (ChlRE), normalized difference vegetation index (NDVI) and normalized difference red-edge (NDRE) were generated. UASDPs were used to predict above-ground biomass-based fresh yield, biological dry yield, and field water-use efficiency (WUEf). This study highlights the use of bootstrap forest partitioning, and partition rank fraction as methods for selecting among many UASDPs as predictors of WUEf and yield. We also used a weighted geometric mean to integrate various model performance metrics to further refine the UASDPs rankings. UASDPs from mid to late growth stages were better predictors of yield and WUEf than UASDPs from earlier periods. Canopy volume, NDRE and ExG were the best predictors overall and were useful for distinguishing varieties, particularly in water-stress conditions. We expect that the approaches detailed here will improve reliability of discriminating the importance of UAS-acquired data types thereby improving UAS data-processing efficiency.