DeepWheat: Estimating Phenotypic Traits From Images of Crops Using Deep Learning
share
In this paper, we investigate the problem of estimating the phenotypic traits of plants from color images and elevation maps of field plots. We focus on emergence and biomass traits - two important indicators of crop growth and health. We employ a state-of-the-art deconvolutional network for segmentation and convolutional architectures, with residual learning in the final stages, for trait estimation. Our intention was to design estimation architectures that behave like high dimensional nonlinear regression models. To the best of our knowledge, this is the first work on emergence counting and biomass estimation based on deep learning. Evaluation was performed on two different species of wheat, grown in field plots for an experimental plant breeding study. Our framework achieves satisfactory performance with mean and standard deviation of absolute difference of 1.20 and 1.53 counts for emergence and 1.45 and 2.05 for biomass estimation. Our results for counting wheat plants from field images is comparable to the accuracy reported for the similar, but arguably less difficult, task of counting leaves from pictures of rosettes grown in pots. Our results for biomass estimation improve upon all previously proposed approaches in the literature.
READ FULL TEXT