Leaf area index-based phenotypic assessment of sweet potato varieties using UAV multispectral imagery and a hybrid retrieval approach

Abstract

Phenotyping based on the estimation of plant traits such as the leaf area index (LAI) could aid the identification and monitoring of the sweet potato health, growth status and gross primary productivity. Integrating radiative transfer models (RTMs), active learning algorithms and non-parametric regression methods using unmanned aerial vehicle (UAV) multispectral imagery have the potential for accurately estimating LAI across multiple crop varieties at varying growth stages. This study tested the boosted regression trees (BRT) and kernel ridge regression (KRR) for inversion of the PROSAIL RTM to retrieve LAI across 20 sweet potato varieties during peak growth stage. Furthermore, the study attempted to constrain the inversion process by using active learning (AL) techniques which ensured the selection of informative samples from a pool of RTM simulations. Results show that the most accurate LAI retrieval over the heterogeneous sweet potato canopy was achieved by integrating smaller PROSAIL simulations with the random sampling AL and KRR methods. The LAI retrieval accuracy had a coefficient of determination (R2) of 0.52, root mean squared error (RMSE) of 0.88 m2.m-2 and relative RMSE of 12.23 %. However, the BRT performance in-comparison to KRR, captured more spatial variability of observed LAI with a better prediction accuracy across the 20 sweet potato varieties. The hybrid approach developed in this study, show potential for accurate phenotyping of LAI dynamics across multiple sweet potato varieties during a matured growth stage. These findings have significant implications for sweet potato breeding programmes that are critical for developing new cultivars in South Africa.