Predicting sugarcane physiological traits using hyperspectral reflectance

Physiological traits have the potential to accelerate genetic improvement for adaptation to abiotic stresses, resource-use efficiency and yield in general. However, using these traits as selection targets in breeding programs is constrained by current phenotyping approaches that involve destructive, time-consuming and labour-intensive measurements. There is growing interest in developing high-throughput tools and prediction models for precise phenotyping of important physiological traits under field conditions. The aim of this study was to explore the potential of remotely piloted aircraft (RPA)-based canopy hyperspectral reflectance in predicting physiological and biochemical traits in sugarcane. Canopy hyperspectral reflectance at the 400–1700 nm spectral region was collected from 10 genotypes grown under three nitrogen (N) treatments in field conditions. Simultaneously, leaf-level physiological and biochemical traits such as photosynthesis, sucrose and starch content were measured to develop partial least squares (PLS) prediction models. Canonical powered partial least squares (cPPLS) models were able to discriminate the N treatments with high accuracy (R 2 = ~0.8) and genotypes with moderate accuracy (R 2 = ~0.6). Partial least square regression (PLSR) models for predicting physiological, biochemical and yield traits from hyperspectral data had varying degree of accuracy. The prediction accuracy was good for cane yield and sugar yield (R 2 = ~0.5), moderate for leaf sucrose, leaf starch content and gas exchange attributes (R 2 = ~0.2), while, poor for the other traits. It appears that a larger spectral and traits dataset from measurements made under different environmental conditions and crop growth stages is needed to improve the PLS prediction model. Results from this initial proof-of-concept study suggests the effectiveness of hyperspectral sensing for characterising and predicting certain physiological and yield attributes. Validation of these results across seasons and under distinct environmental conditions using diverse genotypes is needed before delivering prediction models for phenotyping sugarcane physiological traits using hyperspectral reflectance. Key words High-throughput phenotyping, hyperspectral, partial least square regression, RPA, UAV, drone