Listening to plants: stress detection as a communication problem
How our ACM NanoCom 2026 paper treats a plant as a transmitter and stress detection as a receiver design problem, and why that framing works.
Plants signal all the time. When a plant is under environmental stress, its electrophysiological activity changes. Our team asked a simple question: if you treat that activity as a message, can you build a receiver that reads it reliably?
Reframing the problem
Plant stress detection is usually framed as a plain classification task. In our ACM NanoCom 2026 paper, we reframed it as receiver-side decoding over a biological communication channel. The plant is the transmitter, its electrophysiological signals carry the information, and our job is to design the receiver.
This framing is useful because communication engineering already has a vocabulary for this problem: channel, interference, integration window, likelihood ratio. Once you describe the problem that way, the design decisions become much clearer.
The receiver
The architecture has four stages:
- Interference suppression to clean the raw recordings
- Finite-horizon temporal integration, so each decision uses a window of signal rather than a single instant
- A structured multi-domain feature embedding that captures statistical, temporal, spectral and multiscale characteristics of the signal
- A histogram-based gradient boosting detector that approximates the likelihood ratio in the embedded space
Results
We evaluated the framework on electrophysiological recordings under multiple stress conditions. It reached a macro-averaged F1 score of 0.9879, a ROC-AUC of 0.9992 and a PR-AUC of 0.9993, with an inference time of about 0.10 ms per segment. That makes it accurate and also light enough for real-time use.
The paper received the Best Paper Award at the 13th ACM International Conference on Nanoscale Computing and Communication. I’m grateful to my teammates and advisors at Politecnico di Milano.
What I took from it
The biggest lesson for me: good results came from a good model of the problem, not from the biggest model. Thinking like a receiver designer gave us a pipeline that is accurate, fast and easy to reason about.
Read the paper: doi.org/10.1145/3818305.3830252