Predicting microgreen harvest day from sensor data: can accumulated lux and temperature integral replace a visual check?

Key takeaways

Accumulated lux expressed as daily light integral (DLI) and temperature integral (the sum of degree-hours above a base temperature) together predicted harvest-ready status within plus or minus 0.5 days in 14 of 18 logged sunflower and pea shoot cycles in Bangkok conditions. Visual checks alone called harvest correctly within the same window in 11 of 18. Plant Mood grow log data, 2024.

Temperature integral is the variable most growers are not tracking and the one that explains the most variance in harvest timing. A cycle running 3 degrees Celsius above the average ambient adds roughly one full day of physiological development per week compared to a cycle at the lower ambient -- which means two cycles seeded on the same day can reach harvest readiness 1.5 to 2 days apart with no other variable changed. Plant Mood grow log data, 2024.

Visual checks fail systematically in two directions: growers harvest too early in the first two weeks of running a new variety, and harvest too late after running it long enough to develop a fixed mental image of what 'ready' looks like. Sensor-derived predictions do not drift with familiarity bias.

Accumulated lux alone is an incomplete predictor because it does not account for the temperature-driven metabolic rate that determines how efficiently the crop converts light into biomass. DLI and temperature integral used together outperform either variable used alone by a meaningful margin in Bangkok conditions where ambient temperature variance across seasons is large.

The sensor-derived prediction does not replace a visual check -- it reframes it. Instead of asking 'does this look ready?' the grower asks 'the model says tomorrow, does the tray agree?' That is a different and more productive question, and it catches the trays where biology and expectation have diverged.

TL;DR

Accumulated lux and temperature integral can predict microgreen harvest day more accurately than unaided visual checks, particularly across Bangkok's seasonal temperature swings. The prediction is not magic -- it is basic plant physiology expressed as logged numbers. The visual check is still necessary, but the sensor data gives it a reference point that removes the most common failure modes: harvesting too early because the tray looks ready, or too late because a fixed mental image has drifted out of sync with the actual crop. Both errors cost yield. The numbers catch them before the cut does.

The two trays seeded on the same day that were not ready on the same day

Two sunflower trays, identical seed batch, identical substrate, seeded the same morning. One sat on the upper shelf of the grow rack, one on the lower shelf. Upper shelf ran about 2.5 degrees Celsius warmer on average due to heat stratification in the grow space -- a difference I had measured and logged but had not thought to factor into harvest timing. By day seven, the upper tray looked textbook ready: cotyledons fully open, colour strong, stems upright. I harvested it. The lower tray looked one day behind, which matched the rough mental model I had for the variety. I scheduled it for the following morning.

When I pulled the accumulated lux and temperature integral logs for both trays after the fact, the upper tray had accumulated 23 percent more degree-hours above the 10-degree base temperature by day seven than the lower tray had by day eight -- one full day later. The lower tray at day eight was physiologically less developed than the upper tray at day seven, despite having had 24 more hours in the ground. I had used calendar days and visual impression as my harvest model and produced two trays at different physiological stages, one of which -- the lower tray -- was slightly past optimal by the time I cut it the following morning because ambient temperature climbed overnight. The temperature integral had predicted this. I had not been reading it.

What the prediction accuracy numbers look like

Sensor-derived harvest prediction (DLI plus temperature integral) matched actual harvest-ready status within plus or minus 0.5 days in 14 of 18 cycles versus 11 of 18 for visual checks alone

Plant Mood grow log data, 2024. Across sunflower and pea shoot cycles at Bangkok ambient conditions. 'Harvest-ready' was defined post-hoc as the point at which fresh weight, stem length, and cotyledon colour all fell within the top quartile of outcomes for the variety and cycle. The three cycles where the sensor prediction missed by more than 0.5 days were all affected by mid-cycle lux disruption -- a shading event or light failure -- that the DLI calculation captured but that was not weighted correctly in the prediction model at that stage of development.

Temperature variance across Bangkok grow cycles contributed an average of 1.8 days of harvest timing variance for sunflower between dry season and wet season cycles

Plant Mood grow log data, 2024. Same seed batch, same substrate, same target lux. Dry season cycles running at average ambient 33 to 34 degrees Celsius reached harvest readiness 1.5 to 2 days earlier than wet season cycles running at 29 to 30 degrees Celsius. A calendar-based harvest schedule -- 'sunflower is ready in 8 to 10 days' -- spans exactly this variance without explaining it. The temperature integral explains it.

DLI variance across grow rack positions in a single grow space reached 22 percent between the highest-lux and lowest-lux shelf positions

Plant Mood grow log data, 2024. Measured with a lux meter at each shelf position across a 4-shelf rack over three days. A grower running four trays on four shelves under the assumption that all shelves receive equivalent light is running four different DLI regimes simultaneously. Harvest timing predicted from a single room-level lux reading will be accurate for the middle shelf and wrong for the top and bottom.

Visual harvest checks called the correct day in only 7 of 14 cycles where the grower was running the same variety for the third or later time

Plant Mood grow log data, 2024. The accuracy rate for visual checks declined with familiarity. Growers who had run a variety multiple times showed stronger anchoring to a fixed visual expectation and were less responsive to cycle-specific variation caused by temperature or DLI differences. New-variety visual checks were slightly more accurate because the grower was paying more attention. Familiarity introduces a systematic bias that sensor data does not have.

Post-harvest fresh weight was 11 percent higher on average in cycles harvested using sensor-derived timing versus cycles harvested using visual timing alone

Plant Mood grow log data, 2024. The mechanism is not that sensor timing is magic -- it is that visual timing skews toward slightly early harvest when the grower is uncertain and slightly late harvest when the grower is confident from familiarity. Both directions cost fresh weight. The sensor prediction reduces the variance in harvest timing, and lower harvest timing variance translates directly to more consistent fresh weight outcomes.

How accumulated lux and temperature integral actually work as predictors

Plants do not experience time in calendar days. They experience accumulated energy inputs: light energy that drives photosynthesis and thermal energy that drives metabolic rate. Daily light integral (DLI) is the total light energy delivered to the canopy over a 24-hour period, expressed in moles of photons per square metre per day. In practice, for growers working with lux meters rather than PAR sensors, DLI is approximated by logging lux at canopy level and integrating over the photoperiod -- a rough conversion but sufficient for within-system prediction where the light source is consistent. Temperature integral -- sometimes called thermal time or growing degree units -- is calculated by summing the degrees above a base temperature (typically 10 degrees Celsius for most leafy crops) for each hour of the growth cycle. A day at 30 degrees Celsius contributes 20 degree-hours above base. A day at 25 degrees Celsius contributes 15 degree-hours. The accumulated temperature integral is a running total of the thermal energy the crop has received, which correlates with metabolic rate, enzyme activity speed, and therefore the pace of cell division and biomass accumulation. The two variables do different things. DLI drives the rate of photosynthesis -- it is the input that determines how much carbon the plant can fix per day. Temperature integral drives the rate at which the plant processes and deploys that carbon -- faster at higher temperatures, slower at lower ones. A high-DLI day at low temperature produces carbon that accumulates slowly. A low-DLI day at high temperature burns reserves faster than it builds them. Used together, the two variables describe the crop's physiological budget in a way that neither describes alone.

In practical terms: a sunflower tray at Bangkok dry season ambient (33 degrees Celsius, 14-hour photoperiod, 11,000 lux at canopy) accumulates roughly 180 lux-hours per day and 23 degree-hours above base. A wet season tray at the same lux target but 29 degrees Celsius ambient accumulates the same 180 lux-hours but only 19 degree-hours above base. The dry season tray is running 21 percent more thermal energy per day. Over an 8-day cycle, that is a difference of 32 cumulative degree-hours -- enough to shift the harvest point by 1.5 to 2 days. The grower looking at both trays on day eight and expecting them to be equivalent is looking at crops with meaningfully different physiological ages. Plant Mood grow log data, 2024.

The relationship between temperature integral and development rate is not linear across the full temperature range. Below the base temperature, development effectively stops. Above an upper threshold -- roughly 35 to 38 degrees Celsius for most leafy microgreen varieties -- heat stress begins to counteract the metabolic acceleration and development slows or reverses. Bangkok conditions in dry season regularly push ambient close to this upper threshold, which means the temperature integral calculation should be capped at a maximum effective temperature rather than summed without limit. A cycle running at 36 degrees Celsius ambient is not accumulating more useful thermal time than a cycle at 33 degrees Celsius -- it may be accumulating less, because heat stress is partially negating the metabolic benefit. Growers using temperature integral as a prediction tool in Bangkok dry season should treat any ambient reading above 34 degrees Celsius as a heat-stress flag rather than as a higher-value DLI input.

What the logged cycles showed about prediction accuracy

Comparing variance in harvest timing across 18 logged cycles, DLI variance between cycles was relatively contained -- the target lux was held reasonably consistent and the photoperiod was fixed. Temperature variance between seasons was substantially larger, with dry season cycles accumulating 30 to 40 percent more degree-hours per day than wet season cycles at the same growth stage. When the prediction model was run using DLI alone, residual timing error averaged 1.2 days. When temperature integral was added, residual error dropped to 0.4 days. The temperature variable was carrying most of the predictive weight that the DLI variable missed.

The temperature integral explained more harvest timing variance than DLI across Bangkok seasonal cycles

Growers who invest in lux measurement but not temperature logging are instrumenting the less important variable for harvest prediction in Bangkok conditions. Ambient temperature is the variable with the highest seasonal swing and therefore the highest contribution to timing variance. A cheap temperature logger is more useful for harvest prediction than a more accurate lux meter, if only one additional sensor is being added to the system.

In dry season cycles running at 33 to 34 degrees Celsius, visual checks called harvest readiness on average 0.7 days before the sensor-derived prediction. Post-harvest fresh weight confirmed that the sensor prediction was closer to optimal in 5 of 7 of these cycles -- the visually triggered harvest was slightly early, cutting before the final biomass accumulation window had closed. The mechanism is that heat-accelerated growth produces a tray that looks visually mature -- full cotyledon spread, strong colour -- before it has finished accumulating fresh weight, because cell expansion driven by thermal time runs ahead of the cell division and filling that determines fresh weight.

Visual checks showed systematic early-harvest bias in cycles running warmer than the grower's reference temperature

In hot cycles, the tray looks done before it is done. This is the failure mode that costs the most fresh weight and is the least intuitive to catch because the tray genuinely looks good. A sensor prediction that says wait another 12 hours when the visual says harvest now is delivering real information, not contradicting a correct observation.

Across cycles where multiple trays were run simultaneously on different shelves, the upper shelf consistently accumulated 18 to 22 percent more daily lux-hours than the lower shelf due to inverse square law light falloff and heat stratification contributing a higher temperature integral. The trays on the upper shelf reached sensor-defined harvest readiness an average of 0.9 days earlier than the lower shelf trays. A single harvest schedule applied to all four shelves was wrong for at least two of them in every cycle. Rotating trays between shelf positions at day three or four equalises accumulated DLI and temperature integral across the cycle, producing more uniform harvest timing across the rack.

Rack position introduced a consistent DLI gradient that the harvest schedule did not account for

A four-shelf grow rack without tray rotation is not one grow system -- it is four grow systems with different environmental inputs producing crops at different physiological stages. Treating them as equivalent and harvesting on a single schedule guarantees that some trays are cut early and some late every cycle.

Plant Mood grow log data, 2024

The grower who looks most carefully makes the most systematic errors

Experience with a variety builds a visual reference for what that variety looks like at harvest under the conditions in which it was learned. The problem is that conditions are not constant. Bangkok temperature swings 4 to 5 degrees Celsius between dry and wet season. DLI changes with fixture age, shelf position drift, and any change to the light schedule. When the conditions under which the variety is currently growing differ from the conditions under which the mental reference image was formed, the experienced grower applies a calibrated eye to a miscalibrated target. They see 'ready' when the tray matches the image, not when the tray has actually reached the physiological state the image was supposed to represent. Sensor data does not have a memory of previous cycles. It reads the current cycle's accumulated inputs and makes a prediction based on those inputs alone. That is a feature in a variable environment, not a limitation.

The clearest evidence for this in the grow logs is the fresh weight data across repeat cycles. In the first three cycles on any new variety, visual harvest accuracy was higher than average -- the grower was paying close attention precisely because the variety was unfamiliar and the mental model was not yet anchored. From cycle four onward, fresh weight variance increased even as the grower reported higher confidence in their harvest calls. Confidence and accuracy diverged. The sensor prediction accuracy held flat across all cycles because it had no familiarity bias to drift. Plant Mood grow log data, 2024.

An experienced grower with many cycles on the same variety will make better harvest timing decisions than a sensor-derived prediction, because they have built a detailed mental model of what the crop looks like at each stage.

The harvest timing complaints that sensor data resolves

My harvests are inconsistent even though I am using the same seed batch and the same grow protocol. Some trays yield more than others for no obvious reason.

Harvest timing variance is almost always the culprit for yield inconsistency when inputs are otherwise fixed. Trays cut 12 to 18 hours early lose the final biomass accumulation window. Trays cut 12 to 18 hours late begin early senescence that reduces shelf life and sometimes fresh weight as stored reserves are mobilised. The inputs being the same does not mean the harvest point is the same -- temperature variation across the cycle shifts the optimal window even when seeding date, lux, and substrate are identical. Logging DLI and temperature integral gives each tray its own predicted harvest window rather than applying a fixed schedule to variable conditions.

My sunflower trays in summer are always slightly past their best when I cut them even though I follow my usual timing.

Dry season ambient temperature is accelerating the physiological development of the crop faster than the calendar-based timing accounts for. A tray scheduled for day eight based on experience from a wet season cycle is already at the physiological equivalent of day nine or ten in a dry season grow space running 4 degrees Celsius warmer. The tray has not failed -- it has matured faster than the schedule anticipated. Running the temperature integral across the cycle would show the cumulative degree-hour surplus that explains exactly how many hours early the harvest window opened.

I can't get consistent results across my four-shelf rack. Top shelf trays are always slightly different from bottom shelf trays even when I seed them on the same day.

The shelves are running different DLI and temperature integral regimes. Heat stratification in a non-climate-controlled Bangkok grow space commonly produces 2 to 3 degrees Celsius of vertical temperature gradient across a four-shelf rack. Combined with inverse square law light falloff across shelf positions, the top and bottom shelves may differ by 20 to 25 percent in both DLI and temperature integral per day. Rotating trays at mid-cycle equalises accumulated inputs across the rack. Short of rotation, logging each shelf position separately and applying shelf-specific harvest predictions is the alternative.

Sometimes the tray looks exactly right but after cutting, shelf life is short -- two days instead of four. I don't know what I am missing visually.

Short post-harvest shelf life usually indicates the crop was harvested at peak visual quality but past peak physiological quality -- the cotyledons are fully expanded and the colour is strong, but the cellular turgor and stored nutrient reserves are already beginning to decline. This happens when heat has driven rapid visual maturation ahead of metabolic completion. The temperature integral in these cycles is typically running 15 to 20 percent above the baseline for that variety, signalling that the crop matured faster than usual. Harvesting 12 hours earlier than visual signals suggest -- which feels counterintuitive -- recovers shelf life in these cycles. The sensor data makes this call; the visual cannot.

What the visual check cannot see

A tray can look harvest-ready -- full cotyledon expansion, correct colour, upright stems -- while being at different physiological ages depending on the thermal and light history of that specific cycle. Visual maturity is a snapshot. Physiological age is an integral. A crop that reached visual maturity quickly because of high temperature has done so by burning metabolic reserves faster, not by achieving more development. The post-harvest shelf life and nutrient density of that crop are lower than a slower-developed crop that looks identical at the point of cut. The temperature integral distinguishes these two trays. The visual check does not.

Physiological age versus visual maturity

The optimal harvest window is not a day -- it is a period of roughly 4 to 8 hours within a day, typically in the morning after the lights-on transpiration has refreshed stem turgor and before midday heat begins to reduce it. A visual check at the wrong time of day -- late afternoon when heat has softened the stems and temporarily reduced turgor -- will read a tray as less ready than it is. A visual check in the morning on a tray that hit thermal maturity overnight will read it as right on schedule when it was actually past the window by several hours. The sensor prediction, because it is based on accumulated historical inputs rather than current visual state, is not affected by the time of day at which the check is made.

Intra-day variation in harvest window

Looking at a tray and deciding it is not ready yet is not a neutral act. Every day the tray stays in the grow space is a day of substrate moisture management, airflow management, and mold risk in Bangkok conditions. The grower who adds a day to the harvest schedule because the visual does not yet match the expected image may be adding a day of mold risk to a tray that the sensor data would have called ready. The cost of harvesting one day early -- slightly lower fresh weight -- is almost always lower than the cost of a mold event on a tray held one day too long. The sensor prediction gives a basis for making this tradeoff explicitly rather than absorbing it as a random loss.

The cost of the visual check itself

A grow rack running sunflower, pea shoots, and radish simultaneously is running three crops with different DLI and temperature integral requirements for harvest readiness. Sunflower typically reaches target accumulated DLI in 7 to 9 days at Bangkok lux levels. Radish reaches it in 5 to 7 days. Pea shoots in 9 to 12 days. A visual check calibrated for one variety applied to a mixed rack will produce systematic timing errors on the others. The sensor prediction requires variety-specific calibration but, once built, applies correctly across the rack without the variety-confusion errors that a single visual standard produces.

Variety-to-variety DLI requirement differences that accumulate across a mixed rack

Building a sensor-derived harvest prediction for Bangkok conditions

Geert Warmenbol

Instrument each shelf position separately, not the room. Mount a temperature sensor and a lux meter at canopy level on each shelf and log at 30-minute intervals minimum. Room-level readings average across the gradient that is the actual source of harvest timing variance. A single room sensor is accurate for nothing and misleading for everything. If the hardware budget allows only one sensor placement, choose the shelf that runs the most trays and accept that the prediction will be approximate for the others until each position is instrumented.

A per-shelf DLI and temperature integral log that allows the variance between shelf positions to be quantified within the first two cycles of use. In most four-shelf Bangkok grow racks, the top-to-bottom DLI and temperature integral variance will exceed 20 percent, which is the threshold at which applying a single harvest schedule to all shelves produces reliably inconsistent outcomes.

Build the prediction model from your own data, not from published tables. For each variety, run three cycles where both the sensor-derived prediction and the visual check are logged, the actual harvest is made at the visual trigger point, and fresh weight is recorded. After three cycles, compare the sensor-accumulated DLI and degree-hours at the harvest point across all three cycles. The number that falls within the highest fresh weight quartile is the target threshold for that variety in Bangkok conditions. Temperate-climate DLI tables are starting points, not answers.

After three calibration cycles per variety, a variety-specific DLI and degree-hour threshold that predicts the optimal harvest window within plus or minus 0.5 days in subsequent cycles running under similar conditions. The calibration carries across seasons once the temperature integral component is included, because the model adjusts automatically for ambient temperature differences.

Set a daily degree-hour cap at 24 degree-hours above the 10-degree base temperature -- equivalent to 34 degrees Celsius ambient -- when accumulating the temperature integral. Any ambient reading above 34 degrees Celsius should be logged but capped at 24 degree-hours for that hour in the accumulation model. This corrects for the non-linear relationship between temperature and development rate at the upper end of the Bangkok ambient range, where heat stress begins to reduce development pace rather than accelerate it. Without the cap, the model will predict harvest earlier than the crop actually arrives in hot cycles.

Improved prediction accuracy during Bangkok dry season cycles where ambient regularly exceeds 33 to 34 degrees Celsius. In uncapped models tested across four dry season cycles, prediction error averaged 0.9 days early. In capped models across the same cycles, prediction error dropped to 0.3 days.

Implement tray rotation at mid-cycle -- day three or day four for most varieties -- as part of the harvest prediction protocol rather than as an optional quality step. Rotation equalises accumulated DLI and temperature integral across shelf positions, which means the harvest prediction applies to all trays on a seeding date rather than requiring per-shelf adjustments. It also eliminates the need to explain to anyone helping with harvests why the top shelf is ready today and the bottom shelf is ready tomorrow from the same seeding date.

Reduction in harvest timing spread across a mixed-shelf batch from an average of 1.1 days without rotation to 0.3 days with rotation, based on logged cycles where rotation was introduced as a controlled variable. Harvest operations for a batch become a single-day event rather than a two-day event.

Use the sensor prediction to set the harvest window, then use the visual check to confirm within it. The workflow is not sensor-or-visual -- it is sensor-then-visual. When the accumulated DLI and degree-hours reach 90 percent of the variety-specific threshold, begin daily visual checks. When the threshold is hit, harvest within 12 hours unless the visual check reveals a clear reason to wait -- a quadrant that has not fully expanded, for example, or a stem integrity issue. The sensor prediction narrows the window; the visual check makes the call within it. Neither alone is as accurate as both together.

Fresh weight consistency improvement of 8 to 12 percent across cycles compared to visual-only timing, based on logged data from cycles where the combined protocol was applied versus visual-only cycles run under the same ambient conditions. The improvement is largest in dry season cycles where temperature variance is highest and visual maturity most consistently runs ahead of physiological readiness.

Step 2 -- building the calibration from own data. Growers find a DLI table online, apply it, and wonder why the timing does not match Bangkok reality. The published tables were built in controlled environments at 20 to 24 degrees Celsius with calibrated PAR sensors. They are not wrong for those conditions. They are wrong for a Bangkok apartment grow rack in March. Three cycles of calibration data built from actual logged conditions is worth more than any published table for predicting harvest timing in a specific grow space.