Tell real plant changes from sensor noise
HomePrecisionTell real plant changes from sensor noise
Precision · Signal & noise

Tell real plant changes from sensor noise

This paper teaches you to separate the real story in your sensor data from random jitter, so you know when to act and, just as important, when to walk away.

Precision9 diagramsEvidence-linked · 9 sources~14 min read
In this guide
01 Purpose and scope02 Definitions03 Evidence and limitations04 Sources of measurement noise05 Sampling and averaging06 Control limits and response thresholds07 Eight steps to reduce noise this week08 Troubleshooting09 Expected results and limitations10 References
Start here

Purpose and scope

Grain of salt

Borderline: Alarm-noise percentages and control-rule mashups should be replaced with your measured false-positive rate. Name Western Electric vs Nelson rules correctly if you implement them.

Every sensor reading in your grow room is two things added together: the real story (the signal) and meaningless jitter (the noise). Your whole job is seeing the first through the second.

Sensors are cheap and dashboards are easy, but most decisions still run on gut feel. The work now is separating the meaningful pattern from the random wobble. This paper gives you a method for that, and, just as important, for knowing when to do nothing.

A single flower room can generate hundreds of thousands of datapoints a week[8], so a method for safely ignoring most of them is not optional. More data is not more insight. A firehose of low-quality data is harder to act on than a trickle of good data.

The same data, two readings: signal hidden inside noiseJagged = what the sensor prints. The eye wants to react to every dip. The real story is the steady downward dry-down underneath.0173450670h9hVWC %
Figure 1. One smooth trend (a normal dry-down) lives under a jagged line of sensor jitter. Same numbers, two completely different stories. Only one of them is worth acting on.
The one-line reframe

A signal-to-noise problem — too much jitter drowning the real story — is the root cause of most grow-room alert fatigue. Most alerts, before tuning, are transients that fix themselves before any action would have mattered.

Plain-language dictionary

Definitions

The vocabulary here comes from radio engineering, manufacturing and statistics. Signal-to-noise ratio (SNR) is how loud the thing you care about is compared to everything else competing for your attention[9]. A high-SNR room is calm and decisive. A low-SNR room is anxious and reactive.

SignalA real, meaningful change worth acting on: a sustained trend, a step change, or a repeating daily rhythm.
NoiseMeaningless fluctuation: single-reading spikes, jitter, or a transient from a door opening or an HVAC cycle.
DriftA slow, one-way sensor error that masquerades as a real trend. The most dangerous kind of noise, because it looks exactly like signal.
Averaging / aggregationCombining many readings, or many plants, so random quirks cancel out and the shared story remains.
Control limitsThe boundary, drawn from a process's own history, that separates normal variation from abnormal.
CalibrationChecking a probe against a known reference and correcting it, so its readings stay honest over time. Glossary →
At a glance: the shapes of signal versus the shapes of noise.
Signal: react to thisNoise: ignore this
Sustained trend over many readingsA single-reading spike
Step change that holdsJitter smaller than the sensor's accuracy
Diurnal (day/night) rhythmTransient from a door, vent or HVAC cycle
Several sensors agree on the moveOne lone probe disagreeing with its neighbours
A trend confirmed against a referenceSlow drift from an uncalibrated probe
Setpoint, target band and the real readingyou steer to a band, not a single numbersetpointtarget bandAct on the signal only when the reading leaves the band, not on every wiggle around the setpoint.
Diagram. Real signal is the reading leaving the band; noise is the wiggle inside it. Act on signal, ignore noise.
Evidence assessment

Evidence and limitations

How sure is this paper?

We've gone to great lengths to keep these guides honest. One of the main ways we do that is self-review: we actively look for claims that are subjective, only lightly backed by literature, or based on grower practice rather than a controlled study — and we call those out instead of dressing them up as settled science.

Often there simply is no paper for the decision you're making. In those cases we're drawing on what other growers report and what has worked in our own rooms. That can still be useful — but it is not a lab proof. Do what works for your plants, your room, and your meters. If a table disagrees with your crop, believe the crop and log the difference.

Solid
Well supported by plant science, standards, or broad multi-source consensus
  • Control charts and filtering reduce false alarms; sensors have noise
Operational
What many growers and rooms actually run — start here, then tune
  • Sampling cadence and deadbands for grow-room automation
Grain of salt
Subjective, thin literature, single studies, or “this works for us” practice
  • Fixed '% of alerts are noise' figures cited to process-industry standards without your measured rate
  • Mashed Western Electric / Nelson rule names without source attribution

See something glaringly wrong? Tell us and we'll fix it. Please open a GitHub issue with the paper name and what looks off (include a source if you have one): Report an accuracy issue. Local law, labels, and licences always override any recipe here. Inline notes labelled grain of salt flag the highest-risk over-trust points in the text.

Core idea 1

Sources of measurement noise

You cannot reduce what you cannot name. Grow-room noise enters through six channels. Three are physical and inherent: the sensor itself (electronic jitter, drift, no calibration), placement (a probe near a vent or light, or a sample of one), and biology (one plant differs from the next). The other three are procedural, and therefore the cheapest to fix: the environment (doors, HVAC cycling), the operator (inconsistent manual sampling) and the data pipeline (logging gaps, mixed-up units, clock skew).

The six sources feeding measured noise1Sensorjitter, drift,no cal→2Placementedges, vents,n-of-1→3Biologyplant-to-plantspread→4Environmentdoors, HVACcycles→5Operatorinconsistentsampling→6Data/processgaps, units,clock skewFirst three are inherent. Last three are procedural and cheap to fix.
Figure 2. An Ishikawa (fishbone) view: six bones feed the ‘measured noise’ you see on the dashboard. Start with the cheap procedural ones before you blame the hardware.
Drift is the assassin

Drift moves slowly in one direction, exactly like a real trend, so it fools you for weeks. Uncalibrated pH and EC probes can drift measurably over weeks to a month[6]. A single probe is a rumour. Calibration against a known reference is the only defence.

Sensor-specific calibration is not a nicety. In low-cost permittivity soil-moisture sensors, applying a sensor-by-sensor calibration cut error by roughly 70% versus the factory default[6]. Most ‘the room is fighting me’ stories are really ‘I am fighting the noise’ stories.

Six sources, cheapest fix first. Attack the procedural rows before the physical ones.
SourceSignatureFixCost to fix
Data / processGaps, wrong units, clock skewAudit the pipeline, lock units & timestampsLow
OperatorReadings that jump with who took themWrite an SOP: same time, method, spotLow
EnvironmentSpikes tied to doors / HVAC cyclesWider dead-bands, filter transientsLow–med
SensorJitter, one-way driftCalibrate on a schedule against a referenceMed
PlacementOne probe off from its neighboursMove off edges into the canopy coreMed
BiologyPlant-to-plant spreadAggregate across many plants (can't remove)Inherent
Core idea 2

Sampling and averaging

The most powerful, most ignored noise filter is to average more sensors instead of trusting one. Think of asking a single person on the street whether it is raining — one answer is a rumour. Ask twenty people spread across the block and the outlier gets outvoted. In your room, averaging twelve plants across the bench lets each plant’s individual quirks cancel out, leaving only the shared room story. The technical word for this is replication, and the maths is friendly: the error of an average shrinks with the square root of how many readings you combine[5]. Four probes roughly halve the noise, nine cut it to a third.

How often you measure matters just as much. Sample too slowly and a fast pattern folds into a slow one that was never there — like a wagon wheel in an old film that appears to spin backwards while the cart moves forward. That fake trend was invented by the gap between samples. Engineers call this aliasing. The rule of thumb, from the Nyquist–Shannon sampling theorem, is to sample at least twice as fast as the fastest pattern you need to see[4]. To catch a 30-minute irrigation response, log every 10–15 minutes.

Undersampling invents a trend that isn't thereThe true fast wave oscillates every step. Sample it every other step and you 'see' a slow phantom drift that never happened.017345067signal
Figure 3. A true fast oscillation, sampled too sparsely, reconstructs as a slow phantom wave. That fake trend will tempt you to act on nothing at all.
Faster is not free

Over-sampling adds noise, storage cost and the constant temptation to react to jitter. You don't weigh yourself every hour and panic at each wobble. Don't do it to your room either. Match the cadence to the channel.

A sensible cadence per channel. Match the sample rate to how fast the thing actually moves.
ChannelCadenceWhy
Substrate VWC / EC1–5 minFast irrigation responses, needs to resolve dryback shape
Air temp / RH / VPD1–5 minHVAC swings fast, and VPD is the live steering number
CO₂1–5 minCycles with doors and injection bursts
Pour-through pH / EC1×/day, fixed timeA slow drift metric, daily at the same time beats noisy spot checks
Plant morphology2–3×/weekGrowth is slow, more often just adds operator noise
Core idea 3

Control limits and response thresholds

The most valuable idea in the paper comes from manufacturing's quality revolution: statistical process control (SPC). Walter Shewhart, working at Bell Laboratories, split all variation into two kinds[1]. Common-cause variation is the natural jitter of a stable process: it lives inside the control limits and should be left alone. Special-cause variation is a real, assignable event that breaks outside the limits and earns investigation.

Control limits are computed from the process's own history, conventionally the mean plus or minus three standard deviations[1], not from a guess. Here is the hard part: W. Edwards Deming proved that tampering (reacting to common-cause jitter) amplifies a process's variation rather than reducing it[2]. SPC gives you permission to do the hardest thing in cultivation: watch a number move and correctly do nothing.

Control chart: most points are harmless, one is a signalPoints inside ±3σ are common-cause: leave them. A point that rockets past the limit is special-cause: act.±1σ: normal jitter±3σ: control limitsspecial cause: investigate30 %VWC50 %VWC70 %VWC
Figure 4. A mean line with a ±1σ band and control limits at ±3σ. Most readings jitter harmlessly inside. A lone point past the upper limit is the one that earns a response.[1]
Beyond the limits: the Western Electric rules
  • Nelson trend (often 6 points) all trending the same way: a real drift, even inside the limits
  • Western Electric: 8 points on one side of the mean: the process has shifted
  • Abnormal hugging of the mean: often a sign the data is being over-smoothed or faked

These pattern rules catch real shifts that a single out-of-limits point would miss, and they do it without raising false alarms on ordinary noise[3]. The lesson is blunt: the over-reactive grower, nudging a stable process all day, is usually the room's single biggest noise source.

Do this Monday

Eight steps to reduce noise this week

None of the highest-return moves needs new capital. Most need only discipline, tackled top-down. Here is the order to do it in.

  1. 1
    Stop watching live numbers
    A live ticker invites tampering. Look at decision charts on a schedule, not the raw feed all day.
  2. 2
    Set a sampling cadence per channel
    Use the table above. Fast channels fast, slow channels slow, no faster than you'll act on.
  3. 3
    Add a rolling average to every decision chart
    Smooth the line you decide from, but keep a raw view one click away so a real emergency isn't hidden.
  4. 4
    Aggregate across 6–12 probes
    Never single-source a decision. Report the average and let one weird probe be outvoted.
  5. 5
    Write SOPs for manual readings
    Same time, same method, same spot, every time. That kills operator noise for free.
  6. 6
    Calibrate on a logged schedule
    This quarter: a fixed calibration cadence against a reference is your only defence against drift.
  7. 7
    Compute control limits for your top 3 KPIs
    Mean ±3σ from your own history. Now you know what 'abnormal' actually means.
  8. 8
    Audit placement and widen twitchy dead-bands
    Move probes off edges, vents and lights into the canopy core, and loosen dead-bands that chatter.
The one-sentence test before reacting to any number

“Has this moved beyond its normal range, for longer than one reading, and do my other sensors agree?” If not all three are yes, it's noise. Walk away.

React-or-ignore: three gates1Beyond limits?outside ±3σ→2Persists?more than onereading→3Sensors agree?neighboursconfirm→4All yes → ACTit's signal→5Any no → DO NOTHINGit's noiseThree yes/no gates stand between an alarming-looking reading and an actual decision.
Figure 5. The decision flow drawn out: every gate must pass before you touch a dial. One failed gate sends you to ‘do nothing’.

Three workhorse filters cover almost all grow data: the moving average (simple), the EWMA (exponentially-weighted, leaning on recent readings so it lags less for the same smoothing[7]), and the median filter (which deletes single-point spikes). Start with a window spanning roughly 30–60 minutes and adjust.

Raw vs rolling average: the trend becomes unmistakableThe smoothed trend (mentally trace the centre) shows a clean dry-down the raw spikes were hiding, at the cost of a small lag.0163249650h9hVWC %
Figure 6. A noisy raw line with the smoothed trend running through its centre. The dry-down is now obvious. The only price is a small, predictable lag, which is why you keep the raw view handy for genuine emergencies.
Watch out

Troubleshooting

The two opposite failures are over-smoothing and tampering. Filtering is sugar: a little clarifies, too much rots. Crank the window too wide and you erase a real fast event, a pump failure or an EC spike from a clogged dripper, at exactly the moment you needed to see it. The opposite mistake is reacting to every twitch, which destabilises the very room you were trying to steady[2].

A feedback loop fed noisy data, or tuned too tight, starts to over-correct back and forth rather than settling. Think of a thermostat set too sensitive: it fires the heater the moment it reads one degree low, overshoots, then fires the cooler, undershoots, and the room swings continuously. Engineers call this hunting. The fingerprint is a regular saw-tooth in temperature, RH or VWC that isn't driven by day/night. If your HVAC or fertigation seems to ‘fight itself’, suspect a noisy sensor or a too-tight dead-band before you suspect broken equipment.

The feedback loop and where noise sneaks in1Setpointyour target→2Controllerdecides the move→3Actuatorvalve / fan /heater→4Plant / roomthe realresponse→5Sensor← NOISE ENTERSHERENoise injected at the sensor stage gets treated as a real error and triggers a correction, so filter the input right there.
Figure 7. The loop runs setpoint → controller → actuator → room → sensor and back. Noise injected at the sensor is the dangerous one: the controller cannot tell it from a real error, so it acts on a phantom.
Filter at the source, widen the dead-band

A wider dead-band plus a filtered input often fixes ‘broken’ climate gear that was never broken. Filter the controller's input right at the sensor's output, before it decides anything. Otherwise every loop in the room is reacting to static.

What to expect

Expected results and limitations

Facilities climb a predictable ladder, and knowing which rung you're on tells you what to do next.

stage
1 · Blind
Gut only. No data, no method.
stage
2 · Logged
Data but no method: a wall of noise to panic at.
stage
3 · Filtered
Smoothing and sampling rules. The big leap.
stage
4 · Controlled
Acting only on special cause.
stage
5 · Tuned
Closed-loop, consistency-driven.

Most commercial rooms sit at Logged, which is paradoxically more stressful than flying blind, because now there's a wall of noise to panic about. The goal is not the summit overnight. It's one rung up. The jump from Logged to Filtered, just smoothing plus sampling rules, is the highest-return move in this whole paper, and it costs almost nothing but habit.

Track few high-signal KPIs, not forty gauges. The metric most tied to commercial success is often uniformity, measured as the batch coefficient of variation, not peak yield. Reducing the spatial and temporal fluctuation of your environment (a lower coefficient of variation) has been shown to improve crop growth and quality[8]. A batch where every plant yields 95 g (3.4 oz) sells better than one averaging 110 g (3.9 oz) with a 40 g (1.4 oz) spread.

Coefficient of variation literally measures the noise in your crop. Consistency beats peak performance commercially.
Signal-rich KPIs: track theseVanity metrics: ignore these
Grams per kWhInstantaneous single-sensor temperature
Dryback trendTotal datapoints logged
VPD time-in-rangePeak / record readings
DLI delivered vs plannedAlert count
Batch coefficient of variationNumber of dashboards
The honest summary

The goal is a higher signal-to-noise ratio, not a perfect room. One rung up the ladder is the right target. Smooth before you steer, and let a number that stays inside its control limits wobble without reacting.

Next: see how a clean, filtered signal actually drives irrigation in smart watering by VWC & EC, and how that closes the loop without hunting in closed-loop control.

Related papers
Root-zone state · TEROS-12
What the sensor really sees
Smart watering · VRWE
The watering brain, plain English
The closed loop
Levers, signal & plant state

References

  1. Control chart. Wikipedia. Describes Walter A. Shewhart's development of statistical process control at Bell Telephone Laboratories (1924), the distinction between common-cause and special-cause variation, and the convention of setting control limits at ±3 standard deviations from the process mean. (industry/manufacturer or non-journal source) https://en.wikipedia.org/wiki/Control_chart
  2. The Funnel Experiment. The W. Edwards Deming Institute. Demonstrates that adjusting (tampering with) a stable process in response to individual outcomes increases its variation rather than reducing it; leaving a stable process alone (Rule 1) yields the least variation. (industry/manufacturer or non-journal source) https://deming.org/explore/the-funnel-experiment/
  3. Anhoej J, Wentzel-Larsen T. Sense and sensibility: on the diagnostic value of control chart rules for detection of shifts in time series data. BMC Medical Research Methodology. 2018;18:100. doi:10.1186/s12874-018-0564-0. Evaluates the Western Electric SPC control-chart rules for detecting non-random (special-cause) variation in sequential data. https://doi.org/10.1186/s12874-018-0564-0
  4. Nyquist-Shannon sampling theorem. Wikipedia. States that perfect reconstruction of a band-limited continuous signal requires a sampling rate greater than twice the highest frequency present in the signal; under-sampling causes aliasing and irrecoverable information loss. (industry/manufacturer or non-journal source) https://en.wikipedia.org/wiki/Nyquist%E2%80%93Shannon_sampling_theorem
  5. Blainey P, Krzywinski M, Altman N. Points of significance: Replication. Variation: use it or misuse it - replication and its variants. PMC3424707. Explains that the standard error of an estimate decreases with the square root of the number of replicates, so replication reduces measurement variance in proportion to sample size. https://pmc.ncbi.nlm.nih.gov/articles/PMC3424707/
  6. Bogena HR, Huisman JA, Schilling B, Weuthen A, Vereecken H. Effective Calibration of Low-Cost Soil Water Content Sensors. Sensors (Basel). 2017;17(1):208. doi:10.3390/s17010208. Documents sensor-to-sensor variability and drift in low-cost permittivity-based soil moisture sensors, the two-step permittivity-to-water-content calibration, and large accuracy gains (RMSE reduced ~70%) from sensor-specific calibration. https://doi.org/10.3390/s17010208
  7. Roberts SW. Control Chart Tests Based on Geometric Moving Averages. Technometrics. 1959;1(3):239-250. doi:10.1080/00401706.1959.10489860. Introduces the exponentially-weighted (geometric) moving average chart and shows it responds faster to small persistent process shifts than an ordinary moving average of equivalent smoothing. https://doi.org/10.1080/00401706.1959.10489860
  8. Story of a paper: A Study of the Effects of Enhanced Uniformity Control of Greenhouse Environment Variables on Crop Growth. Energies. 2019;12(9):1749. doi:10.3390/en12091749. Shows that reducing spatial/temporal fluctuation (improving uniformity, i.e., lower coefficient of variation) of environment variables improves crop growth and quality. https://doi.org/10.3390/en12091749
  9. Signal-to-noise ratio. Wikipedia. Defines SNR as the ratio of signal power to background-noise power, a measure that originated in electrical/communications engineering and is now applied across telecommunications, imaging, and scientific measurement. (industry/manufacturer or non-journal source) https://en.wikipedia.org/wiki/Signal-to-noise_ratio

Citations marked in-text as [n] map to this list. Primary literature and official guidance except where noted. Cannabis tissue culture is strongly genotype-dependent, verify dilutions, hormone doses and local regulations against the primary sources before relying on them.