Bayesian adaptive sampling: A smart approach for affordable germination phenotyping.

Digital phenotyping is rapidly advancing, generating increasing amounts of data, particularly in the case of temporal monitoring. We propose an adaptive sampling method that optimizes sampling, thereby reducing costs associated with data production, processing, and storage. The proposed method is based on Bayesian inference, which utilizes previous measurements, historical data, and an expected model. Five Bayesian methods are assessed in this study: Important sampling (IS), Markov chain Monte-Carlo (MCMC), Gaussian process (GP), Extended Kalman filtering (EKF) and Sampling Importance Resampling particle filtering (SIR-PF). We test these five Bayesian sampling methods for the monitoring of germination rate in terms of compression, distortion and computation cost. The best trade-off is found by the MCMC method, which offers a compression rate of 0.2 with very little distortion. GP offers the most unbiased parameter estimation and the capability to adapt to various germination speeds. It also has reasonable computational times.