Abstract

Hyperspectral imaging is a promising method to predict traits in a high-throughput manner with the potential to unlock quantitative genetic studies. Researchers have successfully modeled physiological traits such as vegetative Nitrogen content, but scope of methodology and lack of truly novel testing data hinder large scale trust in the process. Here, I explore the ability to model leaf Nitrogen content from hyperspectral reflectance data collected with a LeafSpec imaging device on 22 maize hybrids. Three broad strategies based on different input feature sets are undertaken. Strategy one mines data for the most informative hyperspectral channels and then constructs a normalized index similar to NDVI as input features. Strategy two considers all 364 channels of hyperspectral data and makes predictions using various machine learning techniques; partial least squares regression(PLSR), random forest regression, and a feed-forward neural net regression. Strategy three aims to take advantage of the spatial distribution of hyperspectral data on the leaf surface by training a convolutional neural net(CNN). A normalized visual index constructed from bands most correlated with nutrient content out-performed established NDVI. PLSR was the most accurate algorithm, followed by feed-forward neural net and then CNN, based on coefficient of determination score. PLSR is well established as a robust method for hyperspectral prediction which is further evidenced by this study. This is one of the first applications of CNN for hyperspectral data. Despite not being the most accurate algorithm there remains room for hyper-parameter optimization.