Time Series Modeling for Phenotypic Prediction and Phenotype-Genotype Mapping Using Neural Networks

Abstract

Image-based high throughput plant phenotyping refers to the process of computing phenotypes non-destructively by analyzing images of plants captured at regular time intervals. The non-invasive measurements of phenotypes at multiple timestamps during a plant’s life cycle provides the motivation to extend the application of time series modeling in the field of phenomic research to (1) predict phenotypes for missing imaging days or for a time in the future based on analyzing past measurements; (2) predict a derived or composite phenotype from its one or more constituents and (3) bridge the phenotype-genotype gap to contribute in the study of improved crop breeding and understanding the genetic regulation of temporal variation of phenotypes. The paper uses long short-term memory, a variant of recurrent neural networks, for phenotype-genotype mapping, while autoregressive neural networks, autoregressive neural network with exogenous input and non-linear input output neural networks are used for phenotypic prediction. The experimental analyses on the benchmark dataset called Phenoseries dataset show the efficacy and future prospects of this foundational study.