Machine Learning Techniques for Predicting Crop Photosynthetic Capacity from Leaf Reflectance Spectra.

Harnessing natural variation in photosynthetic capacity is a promising route toward yield increases, but physiological phenotyping is still too laborious for large-scale genetic screens. Here, we evaluate the potential of leaf reflectance spectroscopy to predict parameters of photosynthetic capacity in Brassica oleracea and Zea mays, a C3 and a C4 crop, respectively. To this end, we systematically evaluated properties of reflectance spectra and found that they are surprisingly similar over a wide range of species. We assessed the performance of a wide range of machine learning methods and selected recursive feature elimination on untransformed spectra followed by partial least squares regression as the preferred algorithm that yielded the highest predictive power. Learning curves of this algorithm suggest optimal species-specific sample sizes. Using the Brassica relative Moricandia, we evaluated the model transferability between species and found that cross-species performance cannot be predicted from phylogenetic proximity. The final intra-species models predict crop photosynthetic capacity with high accuracy. Based on the estimated model accuracy, we simulated the use of the models in selective breeding experiments, and showed that high-throughput photosynthetic phenotyping using our method has the potential to greatly improve breeding success. Our results indicate that leaf reflectance phenotyping is an efficient method for improving crop photosynthetic capacity.