From telemetry to intelligence: the plant-state dashboard
HomePrecisionFrom telemetry to intelligence: the plant-state dashboard
Precision · Dashboards

From telemetry to intelligence: the plant-state dashboard

A grow-room screen should show what the plant is doing, not a wall of raw sensor numbers. Here is how to design one that does.

Precision11 diagramsOperational guide~13 min read
In this guide
01 What this is, and why it matters02 Accuracy, self-review, and grain-of-salt notes03 Key terms, defined from zero04 Why the sensor dashboard falls short05 Six inversions: from gauge cluster to calm dashboard06 What the system actually infers07 The six-layer pipeline08 The new dashboard surface: four zones09 The adoption path: crawl, walk, run10 Trust, confidence, and failure modes11 Measuring success and realistic expectations
Start here

What this is, and why it matters

Grain of salt

Provisional: On-screen advisories (e.g. tip-burn risk timelines) are product-design examples for operator UX — not a validated prognostic model.

A modern grow room is wired with sensors measuring air temperature, humidity, VPD, CO₂, light, substrate moisture, EC, root-zone temperature, pH and power draw, second by second. The dashboards built to show all of this are walls of live numbers and graphs. They tell you what is happening. They never tell you what it means, what is about to happen, or what to do about it.

This paper makes the case for a different design centre, which we will call Plant-State Intelligence: a screen that reasons about the plant instead of just displaying the room. The target is a ‘calm dashboard’: one that stays quiet most of the time and speaks only when it has something worth saying. A telemetry-dump dashboard forces the human to be the integrator, synthesising fifteen graphs into a judgement in real time, often while tired. A plant-state dashboard does that synthesis for you.

What this paper is, and isn't
  • This is an operational and product-design guide, not a horticulture-science paper. Lead-time examples are illustrative UX, not validated predictions. Most claims here are design opinions backed by worked examples.
  • The aim is a screen that infers the plant's state, predicts trouble days early, and prescribes the next step with its evidence and confidence attached.
  • The one-line thesis: a cockpit full of gauges is not a co-pilot.
Where the thinking happens1Sensorsraw streams offthe room→2Wall of graphs15 live charts,no synthesis→3Tired humanmust integrateit all, in realtimeThe conventional path: the human is the integrator.
Figure 1. The telemetry-dump path (above) leaves all the reasoning to the human. Plant-State Intelligence moves that step into the software: sensors → fusion and inference → one plain-language judgement.
How sure is this?

Accuracy, self-review, and grain-of-salt notes

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
  • Showing plant state beats raw sensor walls for operators
Operational
What many growers and rooms actually run — start here, then tune
  • UX patterns and advisory wording examples
Grain of salt
Subjective, thin literature, single studies, or “this works for us” practice
  • Any on-screen clinical prognosis (e.g. tip burn in 48 h) as a validated model

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.

Vocabulary

Key terms, defined from zero

Here are the words before the argument. Don't memorise them. Each one comes back in context.

TelemetryThe raw measurements streamed off your sensors, second by second: temperature, humidity, moisture and the rest, before anything is done with them.
VPD (vapour pressure deficit)How ‘thirsty’ the air is for moisture, which drives how fast a plant transpires. 1.4–1.6 kPa is fine in late flower but punishing in early veg. The same number means different things at different stages.
Crop steeringDeliberately pushing a plant vegetative (leafy growth) or generative (flower and resin) by controlling irrigation and dryback.
InferenceEstimating something you cannot measure directly, like plant stress, by combining several things you can measure.
Sensor fusionCombining several signals over time into one conclusion. ‘Leaf temp up’ alone is noise. ‘Leaf temp up and transpiration flat and dryback unusually deep’ is a diagnosis.
DrybackHow much the substrate dries between irrigations. Dryback depth and dryback rate are derived crop-steering metrics that matter more than any raw moisture number.
Leading vs lagging indicatorA leading indicator is a precursor that warns early. A lagging indicator is a symptom that confirms damage already happened.
Baseline / trajectoryThe expected envelope for this cultivar, at this stage and point in the photoperiod, learned from your own past runs.
PPFD / DLIPPFD is instantaneous light intensity. DLI is the total daily light delivered. EC is the salt concentration in the feed or root-zone pore water.
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. Telemetry is judged against a target band, not a single number; the dashboard flags only real departures.
A daily water-content cycleSaturate to field capacity, hold it, allow a controlled dryback, repeat. The size of the dryback is the steering lever.working bandtoo dry0255075100offP3 minonP1FCP2offwater content %
Diagram. Water content and dryback, read as plant state over the day.
VPD: the climate sweet spotVapour pressure deficit blends temp and humidity into one number; steer to the band for the stage.too humidclones / vegflowerdrystress0.2 kPa1 kPa1.8 kPa
Diagram. VPD as a steerable climate signal.
PPFD through the day adds up to DLIIntensity (PPFD) times hours is the day's total light: DLI in mol/m2/day. The area under this curve is what the plant actually gets.02505007501000offmiddayoffPPFD umol/m2/s
Diagram. PPFD and DLI as the light signal.
The problem

Why the sensor dashboard falls short

The conventional dashboard rests on one implicit theory: ‘expose every measurement and a skilled grower will know what to do.’ That fails in seven predictable ways. Every sensor measures the plant's surroundings, air, root zone, light, and none measures vigour, stress or transpiration directly. That leaves an inference gap the human must cross unaided. Capacitive moisture probes, for instance, report water content in the substrate, never the plant's own water status[4].

It is also reactive. By the time a line crosses a threshold, salt accumulation or a stalled dryback has been accruing for hours or days. Its static high/low alarms ‘cry wolf’: they fire on transient blips like a door opening or a lights-on spike, so growers learn to ignore them. Alarm-management standards from process industries put the alarm-flood threshold at roughly ten alarms per ten minutes and cap high-priority alarms at about five percent. A grow-room dashboard that buzzes constantly has already lost the operator's trust[5].

Seven ways the sensor dashboard fails11 Shows the roomnot the plant→22 Reactivegraph looks badafter stress→33 No memoryno context orhistory→44 Cries wolfstatic alarms,fatigue→55 One signal at a timeno cross-channelview→66 Cognitive loadas if it were afeature→77 Stops at symptomsnever names thecauseThe summary indictment: we built instruments and called them intelligence.
Figure 2. The seven failure modes of a telemetry-dump dashboard. Each is a place where the human is left doing work the software could do.
The lag between cause and symptomEC creeps up for four days. Tip burn only appears on Day 26.23445Day 22Day 23Day 24Day 25Day 26Pore-water EC
Figure 3. Pore-water EC creeps up for four days while the grower notices nothing, until tip burn appears on Day 26. A single-channel chart shows the cause the whole time, but nobody is watching that one line at that moment. That is the lag a plant-state system is built to close[1].
Single-channel widgets hide the truth

The real story about plant health lives in cross-signal, multivariate patterns: moisture, EC, VPD and transpiration moving together. A wall of single-channel gauges structurally cannot express that pattern, no matter how many you add.

The shift

Six inversions: from gauge cluster to calm dashboard

Plant-State Intelligence inverts six assumptions baked into the sensor dashboard. None of these throws the raw data away. It just moves to the ‘basement,’ still available on drill-down for the expert and the post-mortem.

The six inversions. The hardest shift is the last one: silence becomes the default state.
AxisFrom: gauge clusterTo: calm dashboard
ObjectInstrumentation: show the environmentInference: estimate the plant's state
ReferenceFixed thresholdsLearned baselines per cultivar × stage × photoperiod phase
BreadthOne signal per widgetMulti-input fusion across signals
TimingLagging symptomsLeading precursors
OutputAlert: ‘a number moved’Prescription: action, deadline, consequence
PostureAlways-on wall of graphsException-based, quiet by default
The plant should win the fight for attention

An always-on wall of graphs competes with the plant for the grower's attention, and the plant should win. That is why a prescription replaces a bare alert. It names the action, the deadline, and the consequence of ignoring it. And it is why silence, not a full screen, is the healthy resting state.

Core content

What the system actually infers

The pipeline estimates four (really five) interacting states. Environmental state, temperature, RH, VPD, CO₂, light, is reframed as integrals and rates: VPD-hours accumulated today, DLI to date, not instants. Plant response to VPD is non-linear and cumulative rather than tied to any single reading[3], so the accumulated quantity is the meaningful one. Substrate state adds derived crop-steering metrics: dryback depth and rate, field-capacity recovery, and shot-to-shot moisture response.

The plant physiological state is the whole point. It is not measured but estimated, by fusing the others into a transpiration proxy, a stress index, a vigour/stacking trajectory and a steering-response readout. Operational/equipment state and an optional vision state (canopy cameras) round it out. Sensor health itself is treated as a first-class signal, so the system knows when it is blind.

From measured states to the inferred plant state1EnvironmentalVPD-hours, DLI,CO₂→2Substratedryback depth &rate, recovery→3Equipment + Visionpump/valvehealth, canopy→4Plant state (inferred)transpiration,stress, vigour,steeringresponseOuter measured layers feed inward into the one inferred state at the centre.
Figure 4. Environmental, substrate and equipment/vision states feed inward into the inferred plant physiological state. That centre is what the grower actually cares about, and the only thing no sensor reports.
The same number, raw vs derived. The right column is what a plant-state dashboard shows. The left is in the basement.
Raw valueDerived, meaningful form
WC = 42%Dryback depth 8%, slower than this cultivar's baseline
VPD = 1.5 kPa right nowVPD-hours 18% above the in-range envelope for the day
EC = 5.1 mS/cmPore-water EC rising 4 days straight, tip-burn risk
Leaf temp +0.6°CTranspiration flat despite higher VPD: stomata closing
The output is a short list of named conditions

This layer does not emit fifteen numbers. It emits a short list of named conditions: ‘dryback stalling,’ ‘salt accumulating,’ ‘over-transpiring,’ each with a confidence and an evidence chain. Cameras already on site for security become a horticultural input: canopy colour and uniformity, lights-on wilt, height and stacking over days, early discoloration.

Core content

The six-layer pipeline

The system is a six-layer pipeline that maps cleanly onto a Home Assistant–centred stack. Most operations already have layers 0 and 1 without realising it. The intelligence moves to the dashboard long before the actuation does: autonomous control is earned channel by channel, after advisories prove correct.

  1. 1
    Layer 0: Ingest
    Pull every raw stream onto one shared timebase. Most rooms already do this.
  2. 2
    Layer 1: Derive
    Turn raw into meaningful: VPD, dryback %, DLI, recovery slopes, shot response.
  3. 3
    Layer 2: Baseline
    Build per-cultivar / stage / photoperiod envelopes, seeded from horticultural priors and refined on your own runs.
  4. 4
    Layer 3: Infer
    Fuse everything into named conditions with confidence and evidence: rules plus anomaly detection, optionally an LLM reasoning pass.
  5. 5
    Layer 4: Prescribe
    Map each condition to a concrete action with a deadline.
  6. 6
    Layer 5: Present
    The calm dashboard. Optional gated Layer 5b closes the loop on low-risk, explicitly-licensed actions only.
The six-layer pipeline10 Ingestone timebase→21 Deriveraw → meaningful→32 Baselinelearnedenvelopes→43 Infernamed conditions→54 Prescribeaction +deadline→65 Presentcalm dashboardOptional 5b Closed-Loop branches off Present for low-risk licensed actions only.
Figure 5. The pipeline, layer by layer. Layer 2 baselines can be seeded from published horticultural targets (Athena targets are one example) before you have any history of your own.
Advisory-first is the design, not a limitation

The human-in-the-loop posture is deliberate. An operation can run permanently at ‘advise only’ and capture most of the value. Layer 5b auto-applies only the low-risk, explicitly-licensed actions. Anything irreversible or expensive stays human-approved.

Core content

The new dashboard surface: four zones

What the grower opens has four zones, in priority order, and on a good day, three of them are empty. Zone 1 is the Headline: one line of plant truth in plain language with a status colour, which is 90% of what a busy grower needs 90% of the time. Zone 2 is the Watchlist of things drifting but not yet wrong, the precursors, and it exists precisely to make the next zone rare. Zone 3 is Advisories, the only zone that should ever interrupt, each prescriptive and time-bound. Zone 4 is the Evidence and raw basement, demoted but never deleted.

always shown
Zone 1: Headline
‘Flower Day 24 · Room 3 · On-track. Steering generative as intended. No action needed.’
usually quiet
Zone 2: Watchlist
Drifting but not yet wrong: the precursors. Each item is a sentence, not a graph. Often empty.
rare by design
Zone 3: Advisories
The only zone that interrupts. Prescriptive and time-bound. Expands to its evidence chain.
the basement
Zone 4: Evidence / raw
The old dashboard, demoted. Fused signals, baselines, raw graphs, for drill-down and the post-mortem.

Colour and layout do real work here. A calm dashboard leans on pre-attentive cues, a single status colour, position, one bold line, that the eye reads before conscious attention engages, so the ‘all clear’ state is grasped at a glance[2].

A sample advisory, in full

‘Reduce dryback target 3% in Room 3 (Day 24)… Tip burn likely soon (illustrative) if unaddressed. Confidence: high. [Show evidence]’, and the evidence expands to the fused signals, the baseline it violated, and the historical precedent. The old dashboard was 100% Zone 4. The new one leads with Zones 1–3 and keeps 4 in the basement.

How to

The adoption path: crawl, walk, run

This is not a boil-the-ocean rebuild. Each stage ships value and earns the next, and most of the payoff lands by Stage 3, long before any closed-loop control.

The adoption ladder10 Telemetrytoday's rawgraphs→21 Derivemeaningfulmetrics, loweffort→32 Baselinego quiet, killsalarm fatigue→43 Fusewatchlist +advisories comealive→54 Prescribeattach actions +deadlines→65 Closed-loopopt-in, gatedMost of the payoff lands by Stage 3. Stage 5 is optional.
Figure 6. Six rungs from telemetry to closed-loop. Stage 2, baseline and go quiet, is the single biggest step, because it ends alarm fatigue in one move.
  1. 1
    Pick one room, one pattern
    Choose a single failure pattern (say, stalling dryback) and implement Stages 1–3 for just that pattern in Home Assistant.
  2. 2
    Run it shadow-mode for a cycle
    Run alongside the existing dashboard for a full cycle. Don't act on it yet. Watch whether it would have been right.
  3. 3
    Prove the lead time
    Measure the gap between the advisory firing and when the problem would have become visible. Prove it on one advisory before scaling.
  4. 4
    Scale pattern by pattern
    Add the next failure pattern, then the next room. Stage 4 (prescribe) and Stage 5 (closed-loop) are opt-in, channel by channel.
Stage 4 is a stable, valuable end state

Advisory-first is not a stepping stone you are obligated to leave. An operation can sit at Stage 4 forever and capture most of the value. Stage 5 closed-loop is optional and gated to low-risk, licensed actions only.

Pitfalls

Trust, confidence, and failure modes

An advisory system that is wrong and confident is worse than no system at all. Trust is a balance: you spend it with every wrong call and earn it with every right one, so advisory precision, not raw volume, is what drives action. Five guardrails are non-negotiable.

The five guardrails. Each prevents a specific way an advisory system loses the grower's trust.
GuardrailFailure it preventsMechanism
Cold-start honestlyFalse certainty from one cycleSeed from horticultural priors, widen confidence bands, label outputs ‘still learning’
Cheap to correctResentment at wrong callsEvery advisory is dismissable and markable as a false positive, and the marks tune the baselines
Track precisionSilent quality driftAdvisory precision and false-positive rate are first-class, visible metrics
Human in the loopIrreversible or costly mistakesAnything expensive or irreversible stays human-approved
Never a black boxLoss of trust in the WHATEvery conclusion expands to its evidence chain
Treat the system's own blindness as a signal

A drifted, noisy or flatlined sensor is itself an advisory. For example: ‘EC probe in Room 2 reads implausibly flat: suspect failure, EC-derived advisories paused.’ A grower who can't see why will, correctly, stop trusting the what. The system's job is to make the decision obvious, not to make it alone on anything irreversible or expensive.

Reality check

Measuring success and realistic expectations

If the new dashboard is working, the grower looks at it less, is surprised less, and harvests more consistently. Six run-over-run metrics make that concrete, and for half of them, the success direction is down.

Six KPIs and their target direction (illustrative)Bars show a healthy target profile, not measured data. Direction matters more than level.025507510080Lead time ↑15Surprises ↓85Precision ↑25Dwell time ↓70Decisions/wk ↑20Outcome variance ↓
Figure 7. A healthy target profile across the six KPIs. Lead time, precision and decisions-per-week should be high. Surprises, dashboard dwell time and outcome variance should be low.
  • Lead time: hours or days between an advisory and when the problem would have become visible. The core KPI. The whole point is catching drift before it becomes damage.
  • Surprises: visible damage with no prior advisory. Drive this to zero.
  • Advisory precision: acted-upon advisories over total, plus the false-positive rate.
  • Dashboard dwell time: lower is better. Attention should return to the plants, not the screen.
  • Decisions surfaced per week: the output is decisions, not pageviews.
  • Outcome variance: yield and quality consistency, cycle over cycle.
The honest framing

Most of the payoff lands by Stage 3, and advisory-first may well be the right permanent end state. You are never obligated to chase closed-loop control. The working names (Plant-State Intelligence, ‘calm dashboard’) are explicitly placeholders: substance over branding.

Start small. Build the inference layer that catches drift early, see the signal-and-noise paper for the statistics underneath it, and feed it the derived crop-steering metrics from f2 crop steering. The dashboard is only as good as the states it reasons over.

Related papers
Signal & noise
Real signal vs sensor noise
F2 crop steering
The daily P0-P3 cycle
Root-zone state · TEROS-12
What the sensor really sees

References

  1. Pimentel L, Barrueto F Jr. Statistical process control: separating signal from noise in emergency department operations. Journal of Emergency Medicine. 2015;48(5):628-638. doi:10.1016/j.jemermed.2014.12.019. https://doi.org/10.1016/j.jemermed.2014.12.019
  2. Fusco R, Granata V, Setola SV, et al. Visual Perception and Pre-Attentive Attributes in Oncological Data Visualisation. Bioengineering. 2025;12(7):782. doi:10.3390/bioengineering12070782. https://doi.org/10.3390/bioengineering12070782
  3. Grossiord C, Buckley TN, Cernusak LA, Novick KA, Poulter B, Siegwolf RTW, Sperry JS, McDowell NG. Plant responses to rising vapor pressure deficit (Tansley review). New Phytologist. 2020;226(6):1550-1566. doi:10.1111/nph.16485. https://doi.org/10.1111/nph.16485
  4. Briciu-Burghina C, Zhou J, Ali MI, Regan F. Demonstrating the Potential of a Low-Cost Soil Moisture Sensor Network. Sensors. 2022;22(3):987. doi:10.3390/s22030987. https://doi.org/10.3390/s22030987
  5. Engineering Equipment and Materials Users' Association (EEMUA). EEMUA Publication 191: Alarm Systems - A Guide to Design, Management and Procurement; and ANSI/ISA-18.2, Management of Alarm Systems for the Process Industries. (Industry standards; alarm-flood threshold ~10 alarms/10 min, <=3-4 priorities, <=5% high-priority.) (industry/manufacturer or non-journal source) https://www.exida.com/articles/ALARM-MANAGEMENT-AND-ISA-18-A-JOURNEY-NOT-A-DESTINATION.pdf

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.