Unsupervised machine learning via Hidden Markov Models for accurate clustering of plant stress levels based on imaged chlorophyll fluorescence profiles & their rate of change in time

•Novel unsupervised learning method via Hidden Markov Model to discern plant stress.•Time-varying imaged global vs. local plant signal intensities are utilized.•Low-pass spatial filtering of the plant signal profile can improve identification.•Addition of the rate-of-change-in-time of data profiles can improve identification.•Method for automatic determination of cluster number is presented. As far back as 1931, studies have shown that Chlorophyll a fluorescence (ChlF) is a useful tool in plant stress detection. Early and accurate detection of plant stress is invaluable in enabling appropriate and timely intervention. One of the major limitations of past work on ChIF-based plant stress identification is that, often, only a small number of commonly used inflection points, locally oriented within the ChlF transient curve, are utilized to calculate a single index value. These singular values offer limited insight into stressor level or stressor type. In this work, we present an unsupervised method for plant stress classification (identification) that utilizes global (versus local) time-varying ChlF signal data obtained via plant video imaging. The contributions of this work are multi-fold: (a) We classify via clustering these time-varying-intensity signals using unsupervised learning via Hidden Markov Models (HMMs). We show how: (b) the proposed globally based feature selection of a plant’s entire ChlF signal profile, via low-pass filtering, can, in some scenarios, improve classification accuracy of plant stress; (c) the rate-of-change-in-time of the plant’s ChlF intensity time-varying profile, as an additional global feature selection, can further improve the plant stress classification accuracy in some scenarios; (d) quantification can improve the proposed HMM classification method in certain scenarios; (e) HMMs allow more variability in categorizing data via clustering than other raw data distance based metrics. (f) We explore the ergodic and Bakis models for HMM state transition matrix initialization. (g) In addition we propose a new method for initialization of the HMM state transition matrix: state information based initial probability assignment (SIPA) and compare it to the often used heuristic initial probability assignment (HIPA) method. (h) We show how using the Bayesian Information Criterion (BIC) as a performance metric allows a good state number selection (using the quantified data with the proposed state transition matrix initialization methods)