Three Kingdoms on One Timeline

A single plant signal — and within it who was nearby: nobody, a human, a dog, a sheep. The three kingdoms meet in one bioelectric trace.

What it's about

For four parts you have learned to read single beings — humans, animals, plants, whole groups. This chapter finally lays them side by side and shows the astonishing thing: it was always the same procedure, and always the same signals. Here you make this convergence tangible yourself — in a single case where all three kingdoms come together.

In a study from the COINs seminar, students clipped a home-built sensor to a basil plant and set different guests before its nose one after another: sometimes nobody (baseline), sometimes a human, sometimes a dog (a carnivore), sometimes a sheep (a herbivore). Then they asked: can you guess from the plant signal alone who was there? The answer was a surprisingly clear yes — and exactly this you rebuild.

A little background

What the plant gives away. The study found clear, statistically significant differences (Welch test, p < 0.05). Human and dog made the signal more restless — more variance, more noise. The sheep, by contrast, made the signal energy drop, a kind of falling silent — perhaps, the authors suspect, an old reaction to a herbivore. An XGBoost model reached about 83 per cent hits across all four classes, with F1 values above 90 per cent for human and dog; the sheep stayed hardest.

Again an image from the signal. The procedure is familiar to you: the sensor (an OPA2134 amplifier on an ESP32, 380 Hz) cuts the signal into 3-second pieces. From each piece we pull the same sound features as with the cat in 12.1 — MFCCs, spectral centroid, zero-crossing rate, energy and their rates of change. In short: we listen to the plant as if its signal were a sound.

Extract features and train the model

First the features per 3-second window, then XGBoost — the study's winning model. The full code is on GitHub.

import librosa, numpy as np, glob, os
from xgboost import XGBClassifier
from sklearn.preprocessing import LabelEncoder
from sklearn.model_selection import train_test_split
from sklearn.metrics import classification_report

def features(file):
    y, sr = librosa.load(file)                      # load 3-second window
    mfcc  = librosa.feature.mfcc(y=y, sr=sr, n_mfcc=13)
    delta = librosa.feature.delta(mfcc)             # rate of change
    centroid = librosa.feature.spectral_centroid(y=y, sr=sr)[0]
    zcr      = librosa.feature.zero_crossing_rate(y)[0]
    energy   = librosa.feature.rms(y=y)[0]
    return np.concatenate([
        mfcc.mean(axis=1), mfcc.std(axis=1),        # shape + volatility
        delta.mean(axis=1),                         # how fast the signal changes
        [centroid.mean(), centroid.std(),           # frequency centre
         zcr.mean(),      zcr.std(),                # roughness
         energy.mean(),   energy.std()],            # intensity
    ])

CLASSES = ["baseline", "human", "dog", "sheep"]
X, y = [], []
for c in CLASSES:
    for file in glob.glob(f"data/{c}/*.wav"):
        X.append(features(file)); y.append(c)
X = np.array(X)
print("read:", len(X), "windows")

enc = LabelEncoder().fit(y)
Xtr, Xte, ytr, yte = train_test_split(X, enc.transform(y), test_size=0.2,
                                      random_state=0, stratify=y)
model = XGBClassifier(objective="multi:softprob", num_class=len(enc.classes_),
                      eval_metric="mlogloss")
model.fit(Xtr, ytr)

print("overall hits:", round(model.score(Xte, yte), 2))
print(classification_report(yte, model.predict(Xte), target_names=enc.classes_))

The shared timeline

Now the image that gives the chapter its name. Draw the signal energy window by window over time and colour it by who was there. You see the kingdoms side by side — the calm baseline, the restless ups and downs with human and dog, the drop with the sheep.

import matplotlib.pyplot as plt

colour = {"baseline":"#9db3a3", "human":"#6B7A8F", "dog":"#E0A82E", "sheep":"#B5524C"}
plt.figure(figsize=(10, 3.5))
i = 0
for c in CLASSES:
    for file in sorted(glob.glob(f"data/{c}/*.wav")):
        y, sr = librosa.load(file)
        plt.scatter(i, librosa.feature.rms(y=y)[0].mean(), color=colour[c], s=12)
        i += 1
plt.xlabel("window (in order)"); plt.ylabel("signal energy")
plt.title("Three kingdoms, one plant, one timeline"); plt.tight_layout(); plt.show()

Worksheet

When it sticks

Food for thought

• Look closely: it was always the same grip. A voice, a cat sound, a heartbeat, a plant signal — with each you did the same: recorded, turned into an image, let the same machinery read it. A single procedure, from tomato to stock market.
• And deeper: across all kingdoms the same honest signals appear — rhythm, the ups and downs of tension and calm, the return of fixed patterns. On a shared timeline, human, animal and plant move surprisingly close: not because they are alike, but because they are kin in the way they answer their world.
• Keep the boundary here too: what is measured is a correlation between proximity and signal shape — no proof that the plant "knows" who is there. The astonishing need not fear the sober; it holds up to it.

Extension

• Your own kingdoms. Record your own four-class series with the Biolingo sensor: nobody, you, your pet, a second human. Do you reach even close to the study's separating power?
• Temporal shape instead of means. The study tried an LSTM that reads the raw time series instead of averaged features — with a high hit rate for the sheep. Replace the MFCC means with the raw window sequence and compare.
• Confusion matrix. Draw with sklearn.metrics.confusion_matrix which kingdoms are confused most often. Does the sheep, as in the study, fall most readily into the baseline?