Design of a plant-state dashboard for your grow room
HomePrecisionDesign of a plant-state dashboard for your grow room
Precision · Dashboards

Design of a plant-state dashboard for your grow room

We recommend that a grow-room screen shows the state of the plant, and not many numbers from the sensors. This paper shows how to make a screen that finds drift some days before it becomes damage, shows the cause and gives the next step.

Precision11 diagramsOperation guide~13 min to read
In this paper
01 Purpose and scope02 Definitions03 Evidence and limitations04 Limits of dashboards with only sensor data05 Six changes in the design of a plant-state dashboard06 Plant-state inference model07 Structure of a dashboard with six layers08 Structure of the dashboard and its four zones09 Installation procedure10 Trust, confidence and types of problem11 Expected results and limitations12 References
Start here

Purpose and scope

Weak

Temporary: The advisories on the screen (for example, the time before a tipburn risk) are examples of product design for operators. They are not a model for prediction that has validation.

A grow room has sensors for many values: the air temperature, the humidity, the VPD, the CO₂ and the light. The sensors also measure the moisture of the substrate, the EC, the root-zone temperature, the pH and the electrical power that the equipment uses. The sensors measure these values each second. The dashboards that show all of this have many numbers and graphs. They tell you the values that occur. They do not tell you the effect on the plant, the conditions that will occur next, or the task to do.

This paper shows a different method for the design of a dashboard. The name of this method is Plant-State Intelligence. The method gives a screen that calculates an estimate of the state of the plant, and does not only show the values for the room. The target is a ‘calm dashboard’: a dashboard that gives an advisory only when there is something important. Thus it has no signal for most of the time.

A telemetry-dump dashboard makes the person put the information of fifteen graphs together and make a decision, frequently after a long day of work. A plant-state dashboard does this for you.

Information about this paper
  • This paper is a guide for operations and product design. It is not a paper on horticulture. The examples of lead time show a design for operators. They are predictions that have no validation. We recommend most of the designs in this paper. Examples show these designs.
  • The target is a screen with inference of the state of the plant, prediction of problems and prescription of the next step. The screen gives the prediction some days before the problems occur, with its evidence and confidence.
  • The paper starts from this: a screen full of gauges is not a second person that helps you.
Where you make the decision1Sensorsdata from theroom→2Many graphs15 graphs, nosensor fusion→3Personmust put all thedata togetherfastThe usual method: the person puts the data together.
Figure 1. The telemetry-dump method (above) leaves all the work of the decision to the person. Plant-State Intelligence moves this work to the system: sensors → sensor fusion and inference → one decision that is easy to read.
The terms

Definitions

These terms come before the other sections. It is not necessary that you know the terms at this time. Each term occurs again in the paper.

TelemetryThe measurements that your sensors send each second: temperature, humidity, moisture and other values. The data are as they come from the sensors, before you use them.
VPD (vapor pressure deficit)The quantity of water vapor that you must add to the air to saturate it. This quantity changes the rate of transpiration of a plant. A VPD of 1.4–1.6 kPa is correct in the last stage of flowering. In the first stage of vegetative growth, it causes stress. The same number has a different effect at each stage.
Crop steeringThe control of irrigation and dryback to make a plant vegetative (growth of leaves) or generative (flower and resin).
InferenceAn estimate of a value that you cannot measure directly, for example plant stress. You calculate the estimate from some values that you can measure.
Sensor fusionSome signals that you put together in a period of time to give one result. The signal ‘leaf temperature is higher’ is only noise. The signals ‘leaf temperature is higher and transpiration is flat and dryback is deeper than usual’ together are a diagnosis.
DrybackThe quantity by which the substrate dries between two irrigations. Two values for crop steering are the dryback depth and the dryback rate. You calculate them from the sensor data. They are more important than a moisture number from a sensor.
Leading indicator and lagging indicatorA leading indicator is a precursor that gives a signal before damage occurs. A lagging indicator is a symptom that shows damage that occurred before.
Baseline and trajectoryThe usual range of values for this cultivar, at this stage and at this time in the photoperiod. The range comes from the data of your previous cycles.
PPFD / DLIPPFD (photosynthetic photon flux density) is the light intensity at one time. DLI (daily light integral) is the total light that the plants receive each day. EC is the concentration of salt in the feed or in the pore water of the root zone.
Setpoint, target band and the readingControl the value to a band, not to one numbersetpointtarget bandChange the control only when the reading is not in the band. Do not change it for small changes in the band.
Diagram. Compare telemetry with a target band, not with one value. The dashboard shows a warning only for a deviation that is not noise.
Water-content cycle for one dayFill to field capacity, hold it, make a controlled dryback, and do it again. The dryback size is the steering control.correct bandtoo dry0255075100offP3 minonP1FCP2offwater content %
Diagram. The water content and the dryback show the condition of the plant during the day.
VPD: the correct climate bandVPD is one value for the temperature and the humidity together. Keep it in the band for the stage.wet airclone / vegetativefloweringdrystress0.2 kPa1 kPa1.8 kPa
Diagram. VPD is a climate signal that you can control.
The PPFD during the day gives the DLIThe PPFD multiplied by the hours gives the DLI in mol/m2/day. The DLI is the area below this curve.02505007501000offmiddleoffPPFD umol/m2/s
Diagram. PPFD and DLI are the light signal.
Quality of the information

Evidence and limitations

Three types of information

We want each paper to be correct. Thus we examine each paper for information that is not solid. This information comes from the view of one grower, from a small number of sources, or from grower methods that have no controlled test. We identify this information, and we do not show it as solid.

Frequently there is no paper for the decision that you must make. Then we use the reports of other growers and the methods that work in the rooms that we know. This information can help you, but it is not the result of a laboratory test. Use the methods that work for your plants, for your room and for your meters. If a table does not agree with your crop, use the data from your crop and record the difference.

Solid
Tests of plants, specifications or many sources that agree show that the information is correct
  • A display of the plant state is better for operators than a display of many raw sensor readings.
Grower method
Many growers use this method in their rooms. Start with it, then change it for your room.
  • UX structures and examples of advisories.
Weak
The view of one grower, a small number of sources, one test, or a method that works in one room
  • An estimate on the screen of a problem that will occur (for example, tipburn in 48 h) as a model that tests show is correct.

If you find an error, make a report. We will correct it. Make the report on GitHub, with the name of the paper and the information that is incorrect. If you have a source, include it: Make a report of an accuracy problem. The regulations, labels and licenses of your area always override each procedure in these papers. The label weak in a paper shows the points where the risk of an error is highest.

The problem

Limits of dashboards with only sensor data

The usual dashboard uses one assumption that no person states: ‘show each measurement, and a good grower will know the task to do.’ This assumption causes seven frequent problems. Each sensor measures the environment of the plant (air, root zone, light). No sensor measures vigor, stress or transpiration directly. Thus the person must make the inference without aid. For example, capacitive moisture probes give the water content in the substrate, and not the water condition of the plant[4].

The usual dashboard also shows a problem after it occurs. When a value is more than a threshold, the accumulation of salt, or a dryback that stops, started hours or days before. The alarms have constant high limits and low limits, and they give false alarms. They start for short changes, for example when a door opens or when the lights come on. Thus growers start to ignore the alarms.

The alarm management standards of the process industry set the threshold of an alarm flood at approximately ten alarms in ten minutes. They set a maximum of approximately five percent for the most important alarms. The operator does not have trust in a grow-room dashboard that gives alarms continuously[5].

Seven problems of the sensor dashboard11 Shows the roomnot the plant→22 Slowshows stressafter it occurs→33 No recordno baseline orreference→44 False alarmconstant limits,fatigue→55 One signal at a timeno view of othersignals→66 Too much workthe work is a‘feature’→77 Only symptomsdoes not showthe causeThe result: we made instruments and gave them the name ‘intelligence’.
Figure 2. The seven problems of a telemetry-dump dashboard. In each problem, the person does a task that the system can do.
The interval from cause to symptomEC increases slowly for four days. Tipburn occurs only on Day 26.23445Day 22Day 23Day 24Day 25Day 26Pore-water EC
Figure 3. The pore-water EC increases slowly for four days, and the grower does not find a problem, until tipburn occurs on Day 26. A graph of one signal shows the cause at all times, but no person monitors that one graph at that time. A plant-state system decreases this interval[1].
Gauges for one signal do not show the condition of the plant

Do not use only gauges for one signal. The information on the condition of the plant is in many signals together: moisture, EC, VPD and transpiration change together. Gauges for one signal cannot show these changes, also if you add many gauges.

The change

Six changes in the design of a plant-state dashboard

Plant-State Intelligence changes six assumptions of the sensor dashboard to the opposite. No change removes the sensor data. The data move to a lower position on the screen. You can continue to use the data when you want to examine a problem, also after the problem occurs.

The six changes to the opposite. It is not easy to make the last change: the standard state has no signal.
AxisFrom: group of gaugesTo: calm dashboard
ObjectInstruments: the dashboard shows the environmentInference: the dashboard calculates an estimate of the state of the plant
ReferenceThresholds that do not changeBaselines from your data, for each cultivar, stage and photoperiod phase
InputsOne signal for each gaugeSensor fusion of many signals
TimeLagging indicators (symptoms)Leading indicators (precursors)
OutputAlarm: ‘a number changed’Prescription: task, time limit and result
ModeMany graphs, always onA signal only for unusual values. The standard state has no signal.
The grower must monitor the plant, and not the screen

A screen that always shows many graphs causes the grower to monitor the screen and not the plant. The grower must monitor the plant first. Thus a prescription replaces an alarm that only tells you that a number changed. A prescription gives the task, the time limit and the result if you ignore it. Thus a screen with no signal, and not a full screen, is the standard state when there is no problem.

Primary content

Plant-state inference model

The pipeline calculates estimates of four states that have an effect on each other. If you include the vision state, there are five states. The system shows the environment state (temperature, relative humidity, VPD, CO₂ and light) as totals in a period of time and as rates, and not as values at one time. Examples are the VPD-hours since the start of this day and the DLI to this time.

The effect of VPD on a plant does not change at a constant rate. The effect increases with time and does not come from one reading[3]. Thus the total is the important quantity. The substrate state adds calculated values for crop steering. These calculated values are dryback depth, dryback rate, recovery to field capacity, and the change of the moisture after each shot.

No sensor measures the plant state. The system calculates an estimate of it with sensor fusion of the other states. The estimate has a transpiration proxy, a stress index, a vigor and stacking trajectory, and an output for the steering effect. The operation and equipment state and an optional vision state (canopy cameras) complete the model. The system uses the condition of the sensors as an important signal. Thus the system knows when the sensors give no correct data.

From the measured states to the plant state estimate1EnvironmentVPD-hours, DLI,CO₂→2Substratedryback depth +rate, recovery→3Equipment, visionpump and valvestate, canopy→4Plant state (estimate)transpiration,stress, vigor,effect ofsteeringThe outer measured layers go to the plant state estimate in the middle.
Figure 4. The environment, substrate, equipment and vision states go to the plant state estimate. This estimate is the state that is important to the grower, and no sensor gives it.
The same number as a sensor value and as a calculated value. The values on the right are the information that a plant-state dashboard shows. The values on the left are at a lower position on the screen.
Value from the sensorCalculated value
WC = 42%Dryback depth 8%, slower than the baseline for this cultivar
VPD = 1.5 kPa at this timeVPD-hours 18% more than the baseline range for the day
EC = 5.1 mS/cmPore-water EC increased on 4 days in sequence: risk of tipburn
Leaf temperature +0.6°CTranspiration does not change, but the VPD is higher: the stomata close
The output is a short list of named conditions

This layer does not send fifteen numbers. It sends a short list of named conditions: ‘a dryback that stops’, ‘salt accumulation’, ‘high transpiration’. Each condition has a confidence and an evidence chain. Security cameras at the site can supply data for horticulture. They show the color and the uniformity of the canopy, and wilt at the start of the light period. They also show the height and the stacking in a period of some days, and a change of color in the first stage.

Primary content

Structure of a dashboard with six layers

The system is a pipeline with six layers. It agrees with a system that has Home Assistant in the middle. Most operations have layers 0 and 1, and the operators do not know it. The dashboard gets the inference a long time before the control of the equipment gets it. Autonomous control comes for one signal at a time, and only after the advisories for the signal are correct.

  1. 1
    Layer 0: Input
    Put all the data on the same timebase. Most rooms do this.
  2. 2
    Layer 1: Values
    Calculate new values from the sensor data: VPD, dryback %, DLI, recovery rates and the effect of each shot.
  3. 3
    Layer 2: Baseline
    Make baseline ranges for each cultivar, stage and photoperiod. Start with known values from horticulture, and make the ranges more accurate with the data of your cycles.
  4. 4
    Layer 3: Inference
    Use sensor fusion to make named conditions, with confidence and evidence. The methods are the conditions that you set, and tests for unusual values. You can also add a step in which an LLM examines the data.
  5. 5
    Layer 4: Prescription
    Connect each condition to a task with a time limit.
  6. 6
    Layer 5: Display
    The calm dashboard. The optional Layer 5b has a gate. It gives closed-loop control only for tasks that have low risk and that have your approval.
The six-layer pipeline10 Inputone timebase→21 Valuesdata → values→32 Baselinebaselines fromdata→43 Inferencenamed conditions→54 Prescriptiontask + timelimit→65 Displaycalm dashboardOptional 5b: closed loop from Display, only for low-risk tasks that have your approval.
Figure 5. The pipeline, layer by layer. You can start the baselines of Layer 2 with the targets for horticulture in papers (the Athena targets are one example), before you have data from your room.
Advisory-first is the design, not a limit

In this design, a person is in the loop. An operation can stay permanently at the ‘advise only’ stage and get most of the results. Layer 5b makes changes automatically only for tasks that have low risk and that have your approval. A person must give approval for each task that has a permanent effect or a high cost.

Primary content

Structure of the dashboard and its four zones

The screen that the grower opens has four zones, and the first zone is the most important. When the condition of the plant is correct, three of the zones are empty. Zone 1 is the Headline: one short sentence that is easy to read. It tells the condition of the plant, and it has a color that shows the condition. This sentence gives 90% of the information that a grower with much work must have, for 90% of the time.

Zone 2 is the Watchlist, with the items that have drift but are not incorrect at this time (the precursors). The Watchlist makes advisories in the next zone not frequent. Zone 3 is Advisories, the only zone that sends a signal to the grower. Each advisory has a prescription with a time limit. Zone 4 is the zone for evidence and sensor data. It has a low position, but you do not remove it.

always shown
Zone 1: Headline
‘Flower Day 24 · Room 3 · Correct. The steering is generative, the same as the target. No task is necessary.’
usually empty
Zone 2: Watchlist
Items with drift that are not incorrect at this time: the precursors. Each item is a sentence, and not a graph. Frequently empty.
not frequent, by design
Zone 3: Advisories
The only zone that sends a signal to the grower. Each advisory has a prescription with a time limit. It opens to show the evidence chain.
the lowest position
Zone 4: Evidence and sensor data
The previous dashboard, at a lower position. It has the signals after sensor fusion, baselines and graphs of sensor data. You use them to examine a problem, also after the problem occurs.

Color and layout are important here. You see the color, the position on the screen and the headline before you read the sentence. The name for this effect is pre-attentive processing[2]. It is the group of visual properties that you receive automatically, before you select an item to examine. Thus you can see the ‘no problem’ state immediately, and you do not put the information from five panels together.

A full example of an advisory

‘Decrease the dryback target by 3% in Room 3 (Day 24)… Tipburn is possible soon (example) without this change. Confidence: high. [Show evidence]’. The evidence opens to show the signals after sensor fusion, the baseline range and the values that are not in it, and a previous example. The previous dashboard was 100% Zone 4. The new dashboard starts with Zones 1–3 and keeps Zone 4 at a low position.

How to

Installation procedure

It is not necessary to assemble all the system again. Each stage gives results and makes the next stage possible. Most of the results occur by Stage 3, a long time before closed-loop control.

The six stages of use10 Telemetrythe graphs youhave→21 Valuescalculatedvalues, smallwork→32 Baselineno false alarms,no alarm fatigue→43 Fusionwatchlist +advisoriesoperate→54 Prescriptionadd tasks + timelimits→65 Closed loopgate, optionalMost of the results occur by Stage 3. Stage 5 is optional.
Figure 6. Six stages from telemetry to closed loop. Stage 2 (the baseline, with no false alarms) is the largest step, because it stops alarm fatigue in one step.
  1. 1
    Select one room and one type of problem
    Select one type of problem (for example, a dryback that stops). Install Stages 1–3 in Home Assistant for only this type of problem.
  2. 2
    Operate it in shadow mode for one cycle
    Operate the system with the dashboard that you have for a full cycle. Do not use its advisories to do tasks. Examine if each advisory is correct.
  3. 3
    Show the lead time
    Measure the time between the advisory and the time at which you can see the problem. Show this on one advisory before you use the system in more rooms.
  4. 4
    Add one type of problem at a time
    Add the next type of problem, then the next room. Stage 4 (prescription) and Stage 5 (closed loop) are optional, for one signal at a time.
Stage 4 is a stable last stage

You do not have to go from the advisory-first method to a different stage. An operation can stay at Stage 4 permanently and get most of the results. Stage 5 (closed loop) is optional, and it has a gate for tasks with low risk that have your approval.

Problems

Trust, confidence and types of problem

An advisory system that is incorrect and has high confidence is worse than no system. Trust is a balance. Each incorrect advisory decreases the trust, and each correct advisory increases it. Thus the precision of the advisories, and not the number of advisories, causes the grower to do a task. Five guardrails are necessary.

The five guardrails. Each guardrail prevents one problem that decreases the trust of the grower in an advisory system.
GuardrailProblem that it preventsMechanism
Show low confidence at cold startToo much confidence from the data of one cycleStart with known values from horticulture. Make the confidence ranges larger. Put the label ‘data not sufficient’ on the outputs.
Easy to correctThe grower does not want the system after incorrect advisoriesThe grower can remove each advisory and identify it as a false positive. The identifications adjust the baselines.
Monitor precisionA drift in quality with no signalThe precision of the advisories and the false-positive rate are important values that the grower can see.
Person in the loopErrors with a permanent effect or a high costA person gives approval for each task with a high cost or a permanent effect
Not a black boxNo trust in the advisoryEach advisory opens to show its evidence chain
A sensor problem is an advisory

Make an advisory for a sensor that has drift, noise or a flat signal. For example: ‘The EC probe in Room 2 gives a flat value, and this value is not possible: the probe can be defective, and the system stops the advisories from EC.’ If a grower cannot see the cause of an advisory, the grower will, correctly, have no trust in the advisory. The system must make the decision easy for the grower to see. A person must make the decision for a task with a permanent effect or a high cost.

The results

Expected results and limitations

When the new dashboard operates correctly, the grower monitors it less, has less damage with no advisory, and gets harvests with more uniformity. Six key performance indicators (KPIs) from one cycle to the next show this. For three of them, the target direction is down.

Six KPIs and their target direction (example)The bars show a target profile, not measured data. The direction is more important than the value.025507510080Lead time ↑15Damage ↓85Precision ↑25Screen time ↓70Decisions ↑20Result variance ↓
Figure 7. A good target profile for the six KPIs. We recommend high values for lead time, precision and the number of decisions each week. We recommend low values for damage with no advisory, screen time and result variance.
  • Lead time: the time (hours or days) between an advisory and the time at which you can see the problem. Lead time is the primary KPI. The system must find drift before it causes damage.
  • Damage with no advisory: damage that you can see, with no advisory before it. Decrease this to zero.
  • Advisory precision: the number of advisories that the grower used for a task, divided by the total number of advisories. Also monitor the false-positive rate.
  • Screen time: the time that the grower monitors the dashboard. A lower value is better. The grower must monitor the plants and not the screen.
  • Decisions each week: the number of decisions that the dashboard gives in one week. The output is decisions, and not views of the screen.
  • Result variance: the variation of yield and quality from one cycle to the next.
Closed-loop control is optional

Most of the results occur by Stage 3. It is possible that advisory-first is the correct permanent last stage. You do not have to use closed-loop control. The names in this paper (Plant-State Intelligence, ‘calm dashboard’) are only temporary names. The content is more important than the name.

Start with a small system. Make the inference layer that finds drift in the first stage. Refer to the signal-and-noise paper for the statistics that it uses. Give the layer the calculated values for crop steering from f2 crop steering. The dashboard is only as good as the states that it uses.

Related papers
Signal and noise
Find the difference between plant changes and sensor noise
F2 crop steering
The P0-P3 cycle for each day
Root-zone condition · TEROS-12
How to read the values from the sensor

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.) (source not from a journal, for example from a manufacturer) https://www.exida.com/articles/ALARM-MANAGEMENT-AND-ISA-18-A-JOURNEY-NOT-A-DESTINATION.pdf

Each citation in the paper, for example [1], refers to an item in this list. Each item is a primary source or an instruction from an authority, unless the item shows a different type. The results of cannabis tissue culture change much with the genotype. Before you use the dilutions, hormone doses and local regulations, make sure that they agree with the primary sources.