Automated Phytosensing: Ozone Exposure Classification Based on Plant Electrical Signals

TL;DR

The paper addresses the lack of automated methods to classify ozone exposure from plant electrophysiology for phytosensing-based urban air quality monitoring. It introduces a workflow combining tsfresh for generic time-series features, Naive AutoML for automated pipeline search, and adapted forward feature selection to identify high-performing models. The approach achieves up to 94.6% accuracy on unseen ivy leaf data and shows generalizability to other species and stimuli, indicating broad applicability. This work supports scalable, low-cost urban monitoring using living plants and provides a reproducible framework for real-time phytosensing deployments.

Abstract

In our project WatchPlant, we propose to use a decentralized network of living plants as air-quality sensors by measuring their electrophysiology to infer the environmental state, also called phytosensing. We conducted in-lab experiments exposing ivy (Hedera helix) plants to ozone, an important pollutant to monitor, and measured their electrophysiological response. However, there is no well established automated way of detecting ozone exposure in plants. We propose a generic automatic toolchain to select a high-performance subset of features and highly accurate models for plant electrophysiology. Our approach derives plant- and stimulus-generic features from the electrophysiological signal using the tsfresh library. Based on these features, we automatically select and optimize machine learning models using AutoML. We use forward feature selection to increase model performance. We show that our approach successfully classifies plant ozone exposure with accuracies of up to 94.6% on unseen data. We also show that our approach can be used for other plant species and stimuli. Our toolchain automates the development of monitoring algorithms for plants as pollutant monitors. Our results help implement significant advancements for phytosensing devices contributing to the development of cost-effective, high-density urban air monitoring systems in the future.

Figures (7)

• Figure 1: Experimental setup to expose an ivy plant to ozone. The ivy plant is within a Faraday cage to minimize external disturbances. We use two PhytoNodes Buss2024 to measure the plant electric differential potential and a RaspberryPi to store the data. Ozone is generated and induced from the top. The ozone concentration is controlled via two ozone sensors (one at the top and one at the bottom).
• Figure 2: Example data of the first experiment with 17 expositions. The solid lines show the mean scaled electric differential potential (EDP) over all expositions of the ivy plant (blue at the leaf, orange at the stem). The shaded areas give the standard deviation. The green time interval shows the ten minute slice selected for calculating the stimulus features, the red time interval shows the ten minute slice selected for calculating the no stimulus features, and the gray time interval indicates the ten minute slices used for background subtraction.
• Figure 3: Each ROC curve is averaged (solid line) over 500 splits using 80%/20% stratified shuffle split and all features of the analysis datasets (blue: leaf, orange: stem, and green: combined). The shaded area corresponds to the standard deviation. The suggested pipeline from Table \ref{['tab:results_AutoML']} is used for each setting.
• Figure 4: Accuracy results for a simple threshold model (blue: leaf, orange: stem, and green: combined). The model is evaluated using all experiment data and the result is averaged over 500 independent runs with a random 80%/20% train/validation split. See our implementation for details.
• Figure 5: Mean ROC AUC score (solid line) for the all feature feature subsets of the (a) leaf, (b) stem, and (c) combined analysis dataset. The best performance is reached using (a) 62 features (AUC ROC score of 0.9901) for the leaf analysis dataset, (b) 69 features (AUC ROC score of 0.9063) for the stem analysis dataset, and (c) 94 features (AUC ROC score of 0.9985) for the combined analysis dataset. Shaded area is the standard deviation.