1. Introduction
Protected cultivation is expanding in hot and humid regions, where greenhouses shelter crops from heavy rain and insects but trap heat [1, 2]. Without active cooling the air in such a house can pass 47 °C and stay above 70 % relative humidity on most days [3], and tomato grown under net screens in the Malaysian lowlands has been exposed to 39.7 °C and 98.9 % relative humidity [4]. Fan-and-pad evaporative cooling is the usual remedy, and it is a resource cost as well as a climate control: in one fan-and-pad house the pads used 8.4 to 14 l of water and the fans 48 to 100 W h of electricity per square metre of floor each day, and both rise with the hours the controller keeps them running [5, 6]. That controller acts on what its sensors report. A temperature sensor that reads 2 °C high would keep the fans and pads running longer than the crop requires, and a humidity sensor that reads 5 % relative humidity too high would shift every humidity-based decision by the same amount, so the quality of the measurement bears directly on the water and energy the house uses.
Low-cost sensor nodes have made dense monitoring of greenhouse climate affordable. A microcontroller such as the ESP32 with digital temperature, humidity and light sensors costs a small fraction of a research-grade station and can publish readings over Wi-Fi to cloud dashboards at intervals of a minute or less [7, 8, 9]. Their accuracy in the field, however, is usually taken from the manufacturer’s datasheet or from a short comparison on the bench, and few low-cost temperature sensors have been calibrated in the field [10], and evaluations in the field have found significant humidity bias even where the correlation with the reference was high [11]. Neither study describes a greenhouse. Unshielded sensors warm in sunlight and read high [12, 13, 14], capacitive humidity elements drift, most at high humidity [15, 16], and inexpensive light sensors saturate in direct sun [17, 18]. Evaluations against reference instruments have shown that a linear correction removes much of the offset of such sensors [19, 20].
Two questions about these corrections are rarely asked. The first is whether a correction holds outside the conditions it was fitted on. Calibrations of low-cost sensors are most often made indoors, and the range they cover changes their outcome [10]. The air in a tropical greenhouse, by contrast, spanned 28 °C–48 °C and 19 %–82 % relative humidity over the test days of this study (Section 3.4), so a controller can spend much of its time in conditions the calibration never saw. The second concerns the transmission interval, which is usually chosen for battery life, bandwidth or storage. Observation guidance treats a 1 min average as most suitable and accepts up to 10 min [21], but whether a 10 min series carries the same information for control as a 1 min series depends on how fast each quantity changes, and light under broken cloud changes within seconds [22]. Comparing intervals recorded on different days, moreover, mixes the effect of the interval with that of the weather.
Herein, we calibrated and validated three low-cost nodes (ESP32 with SHT20 and BH1750 sensors) against reference instruments in a fan-and-pad cooled greenhouse in Lampung, Indonesia, and followed them through 1458 paired readings per quantity at three times of day. We asked whether the corrected sensors agree with the reference within its stated accuracy, whether that agreement depends on the range covered by the calibration, and, by subsampling the 1 min series, how much information a 5 or 10 min interval loses. The greenhouse is one example of a class of settings, evaporatively cooled houses in the humid tropics monitored with low-cost nodes. In this house we measured how, and by how much, a short field calibration fails across a wide diurnal range. The results bear on two questions that other houses can test: how wide a calibration window must be, and which quantities tolerate a sparse transmission interval.
2. Materials and Methods
2.1 Greenhouse and Measurement Points
Measurements were made from March to June 2024 in the arch-type research greenhouse, covered with translucent plastic, on the campus of Institut Teknologi Sumatera (ITERA), Lampung Selatan, Indonesia (5.37° S, 105.31° E); its floor dimensions were not recorded. The greenhouse is cooled by an evaporative pad at one end and exhaust fans at the other; whether the pad and fans ran during each measurement was not recorded. Three identical sensor nodes were placed along the long axis of the house (Fig. 1a): point 1 near the cooling pad, point 2 at the centre, and point 3 near the exhaust fans, at 2.7, 12.7, and 22.06 m from the pad end and 4.1 m from the side wall. Each node was mounted on a vertical pole with its sensors about 1.5 m above the floor and the light sensor facing upwards (Fig. 1c).
2.2 Monitoring System
Each node consisted of an ESP32 microcontroller (NodeMCU), an SHT20 digital sensor for air temperature and relative humidity (RH), and a GY-30 module carrying a BH1750 ambient-light sensor (Table 1). Both sensors shared the I2C bus of the ESP32 (SDA on GPIO 21, SCL on GPIO 22) and were powered at 3.3 V. The sensors and the ESP32 were fixed to one foam disc (Fig. 1b), and the disc was mounted under a conical cover with a louvred cap (Fig. 1c), which shaded the electronics from direct sun but was not ventilated by a fan; the sensor therefore exchanged heat with the air only by natural convection, and it shared that enclosed space with the ESP32. Three time scales are involved in each transmitted value and are kept apart here: the conversion time of the sensors, below 0.1 s for the SHT20 [16] and about 0.12 s for the BH1750 in its high-resolution mode [17]; the rate at which the firmware read the sensors; and the transmission interval of 1, 5 or 10 min. The firmware transmitted a single instantaneous reading taken at transmission time, not an average of readings taken between transmissions; the BH1750 measurement time actually set was not recorded. Readings were published over Wi-Fi to an MQTT broker, passed through a REST interface to a cloud server, and displayed with their timestamps in the ITERAHERO web application. The transmission interval was set in firmware to 1, 5, and 10 min in the respective performance tests (Section 2.5).
| Instrument | Quantity | Range | Stated accuracy |
|---|---|---|---|
| SHT20 | Air temperature | −40 °C to 125 °C | ± 0.3 °C typical (5 °C–60 °C) [16] |
| Relative humidity | 0 %–100 % | ± 3 % RH typical (20 %–80 % RH) [16] | |
| BH1750 (GY-30) | Illuminance | 1 lx–65535 lx | ± 20 % measurement variation [17] |
| HTC-2 thermohygrometer | Air temperature | −10 °C to 50 °C | ± 1 °C |
| Relative humidity | 10 %–99 % | ± 5 % RH |
|
| UNI-T UT383 luxmeter | Illuminance | 0 lx–199999 lx |
|
2.3 Reference Instruments
The reference for air temperature and relative humidity was the
HTC-2 thermohygrometer, a generic hand-held instrument without a
manufacturer’s datasheet or calibration certificate, and that for
illuminance the UNI-T UT383 luxmeter
(Table 1). Because
neither instrument carried a calibration traceable to a standard,
all results in this paper describe agreement with these reference
instruments, not traceable measurement accuracy. To size the
uncertainty this leaves, the stated accuracy of each instrument was
treated as the half-width of a rectangular distribution and combined
with its display resolution. For the HTC-2 this gives standard
uncertainties of 0.58 °C and 2.9 % RH (expanded,
2.4 Calibration and Validation
At each point, sensor and reference readings were paired every minute from 16:00 to 18:00 local time (UTC+7), giving 121 pairs per quantity and point. The calibration therefore covered the conditions of a single late-afternoon window: reference values of 31.3 °C–39.3 °C, 41 %–65 % RH and 4.1 klx–35.4 klx across the three points. The window was chosen because in the late afternoon all three quantities change slowly and in one direction, temperature and illuminance falling and relative humidity rising, so that one session spans a range of each, and because the readings are less disturbed by direct sun and passing cloud than at midday; only this window was used. Calibration and validation were carried out on different days, whose dates were not recorded; the validation day was warmer (reference up to 43.7 °C) and spanned 34 %–69 % RH, so the two data sets did not share the same conditions. The calibration data of each quantity and point were fitted by ordinary least squares as
2.5 Performance Tests
With the corrected firmware, sensor and reference readings were recorded at the three points in three one-hour windows (07:00–08:00, 12:00–13:00 and 16:00–17:00), at transmission intervals of 1, 5 and 10 min, which yield 61, 13 and 7 readings per window. Each combination of window and interval was repeated on two days. Each interval was run on its own pair of days, so that differences between intervals include day-to-day differences in weather; the dates were not recorded.
Reference readings were taken at the three points in the same
minutes, with one set of reference instruments per point. The tests
yielded 486 reading pairs per point and quantity (1458 per
quantity). The display resolution of the reference instruments is
0.1 °C, 1 % RH, and 1 or 10 lx below and above 10 klx, respectively
(Table 1). In 43 of
the 162 one-hour windows (54 per quantity: 3 points
Because each interval was run on its own days, the effect of the interval was assessed on the 1 min windows only, by subsampling each 61-reading series at every fifth and every tenth minute (13 and 7 readings). The hourly mean, maximum, minimum and standard deviation of the subsampled series were compared with those of the full series, which holds the weather constant between intervals.
2.6 Data Analysis
For
Consecutive one-minute readings are autocorrelated, so their
scatter overstates the information they carry. The 95 % confidence
intervals (CI) for bias, MAE, RMSE and MAPE were therefore obtained
by block bootstrap with 4000 resamples [25]:
for the performance tests, whole one-hour windows were resampled (18
per point and quantity), and for the two-hour (121-reading)
calibration and validation series, moving blocks of 10 min. The
coefficient of determination describes linearity only; a sensor with
a constant offset can have
To separate a humidity offset from sensor heating, the uncorrected calibration readings were converted to vapour pressure and dew point with the Magnus approximation [27], and the relative humidity implied by the reference vapour pressure at the sensor’s own temperature reading was compared with the sensor’s humidity reading.
The agreement criterion was applied at two levels. For a data set, a corrected sensor was taken to agree with the reference when its bias and RMSE fell within the stated accuracy of the reference instrument (Table 1; temperature and humidity only, illuminance being compared by MAPE), since a difference smaller than that accuracy cannot be resolved against the reference. For individual readings, the share whose difference lay within ± 1 °C or ± 5 % RH was reported for each class relative to the calibrated range (below, within, above); no such share is given for illuminance, whose reference tolerance varies with the reading.
The number of reading pairs entering each stage was: calibration
and validation, 121 per point and quantity (363 per quantity in each
stage); performance tests, 486 per point and quantity, of which 366,
78 and 42 at the 1, 5 and 10 min intervals (1458 per quantity, 4374
in total); and the subsampling comparison, 18 one-hour windows of
the 1 min series per quantity (3 points
3. Results and Discussion
3.1 Calibration and Validation
Before correction, every SHT20 read warmer and wetter than the reference (Fig. 2a,b; Table 2). The temperature bias was 1.51 °C–2.43 °C across the three points, and the slope of Eq. (1) was 1.34–1.65, so the offset grew with temperature: at point 1 the sensor read 43.8 °C against 38.7 °C at the start of the calibration window, when the house was warmest, and 32.7 °C against 31.4 °C as it cooled. The relative-humidity bias was 5.2 %–8.5 % RH, above the ± 5 % RH accuracy of the reference at all three points. Neither offset was traced to a cause in this design, and the explanations below are candidates. A temperature offset that grows with temperature is what a sensor body warmed above the surrounding air would show: poorly shielded sensors have read 0.8 to 3.0 °C warm in the field, with the bias strongest at the highest temperatures [12], and a Sensirion sensor mounted on the circuit board of a wireless node carried an error that followed solar radiation [20]. Both sources of heat were present here: the SHT20 sat on the same disc as the ESP32, a few centimetres from it, under a cover that shaded it but was not ventilated (Fig. 1b,c). Because the two sources acted together in every reading, and neither solar radiation nor board temperature was measured, their shares cannot be separated, and a contribution of the reference itself cannot be excluded. For the humidity offset one inference does follow from the data. Converted to vapour pressure with the Magnus approximation [27], the uncorrected sensors held 22–27 % more water vapour than the reference (dew point 3.4–4.0 °C higher), whereas a sensor that was only warmer than the air would have read 4.1–6.4 % RH below the reference, because heating biases a capacitive element dry [28, 16]. Heating of the sensor therefore does not explain the humidity offset; the offset lies in the humidity element, in the reference, or in both, and the data cannot decide between them. A drift of the element is one possibility: this sensor is specified to develop a reversible positive offset of 3 % RH after 60 h above 80 % RH [16], and in-service drift of up to 5 % RH has been observed at the high end of the range [15]. A reference error of the same sign is also possible, given an HTC-2 standard uncertainty of 2.9 % RH (Section 2.3). The BH1750 read within about 10 % of the luxmeter without correction (MAPE 8.8 %–10.6 %), with a slope of 0.95–1.07.
| Before correction (calibration) | After correction (validation) | |||||||
|---|---|---|---|---|---|---|---|---|
| 2-6(lr)7-9 Point | Bias | RMSE | MAPE | Bias | RMSE | MAPE | ||
| Air temperature (°C) | ||||||||
| 1 (pad) | 1.65 | 0.997 | 2.43 | 2.75 | 7.0 | 0.28 | 0.34 | 0.8 |
| 2 (centre) | 1.40 | 0.995 | 1.64 | 1.80 | 4.7 |
|
0.75 | 1.0 |
| 3 (fan) | 1.34 | 0.991 | 1.51 | 1.73 | 4.3 | 0.34 | 0.54 | 1.0 |
| Relative humidity (% RH) | ||||||||
| 1 (pad) | 0.96 | 0.994 | 5.2 | 5.3 | 9.6 |
|
1.1 | 1.8 |
| 2 (centre) | 0.86 | 0.991 | 8.5 | 8.5 | 16.2 | 2.2 | 2.6 | 4.9 |
| 3 (fan) | 0.90 | 0.995 | 6.7 | 6.7 | 12.3 | 0.1 | 1.1 | 1.8 |
| Illuminance (lx) | ||||||||
| 1 (pad) | 0.95 | 0.974 |
|
1 814 | 10.6 |
|
296 | 7.1 |
| 2 (centre) | 1.07 | 0.986 |
|
839 | 8.8 | 386 | 615 | 6.9 |
| 3 (fan) | 1.04 | 0.975 |
|
1 145 | 10.6 |
|
487 | 6.0 |
After the correction was applied in firmware, the validation readings fell on the 1:1 line (Fig. 2d–f). The temperature RMSE dropped from 1.73 °C–2.75 °C to 0.34 °C–0.75 °C and the humidity RMSE from 5.3 %–8.5 % RH to 1.1 %–2.6 % RH, both inside the accuracy of the reference, while the illuminance MAPE fell from 8.8 %–10.6 % to 6.0 %–7.1 %. The validation temperature RMSE is of the same size as the standard uncertainty of the reference, 0.58 °C, and the humidity RMSE is below its 2.9 % RH (Section 2.3). These values therefore show that the corrected sensors agree with the reference to within the resolution of the comparison; they do not establish an accuracy better than ± 1 °C or ± 5 % RH, because the reference itself is specified no tighter. Bland–Altman limits of agreement for the validation, pooled over the three points, were −0.95 °C to 1.22 °C and −2.5 % to 3.8 % RH. The validation also carried the first sign of the range limit discussed below. The validation day was warmer than the calibration day (reference up to 43.7 °C against 39.3 °C), and above 40 °C the corrected point 2 flattened at 38.5 °C–39.6 °C while the reference read 40.1 °C–41.6 °C (Fig. 2d).
3.2 Field Agreement and the Calibrated Range
Over the 486 readings per point of the performance tests, the
corrected SHT20 read 0.73 °C warm on average (RMSE 1.16 °C) and the
humidity sensor read 1.9 % RH dry (RMSE 3.8 % RH), pooled over the
three points (per point in
Table 3). Both summaries
hide a dependence on where the reading lay relative to the
calibrated range. Of the temperature readings, 33 % were above the
39.3 °C ceiling of the calibration, 92 % of them at midday, and 16 %
below its 31.3 °C floor, all in the morning
(Fig. 3a). Within the
calibrated range the mean bias was 0.43 °C (95 % CI 0.22–0.71) and
83 % of readings agreed with the reference to within ± 1 °C. Above
it the bias rose to 1.19 °C (0.75–1.62) and only 38 % of readings
stayed within ± 1 °C; below it the bias was 0.75 °C (0.47–0.98)
with 76 % within tolerance. The Bland–Altman analysis shows the same
asymmetry (Fig. 4a). Within
the range the difference did not depend on temperature (slope
| Subset | Bias | MAE | RMSE | MAPE | |||
|---|---|---|---|---|---|---|---|
| Air temperature (°C) | |||||||
| Point 1 (pad) | 486 | 0.38 | 0.49 | 0.65 | 1.3 | 1.02 | 0.985 |
| Point 2 (centre) | 486 | 0.59 | 0.85 | 1.07 | 2.3 | 0.99 | 0.970 |
| Point 3 (fan) | 486 | 1.22 | 1.28 | 1.58 | 3.3 | 1.08 | 0.972 |
| Below range | 233 | 0.75 | 0.75 | 0.84 | 2.5 | 0.98 | 0.809 |
| Within range | 746 | 0.43 | 0.58 | 0.86 | 1.6 | 0.86 | 0.853 |
| Above range | 479 | 1.19 | 1.39 | 1.62 | 3.2 | 1.10 | 0.817 |
| Relative humidity (% RH) | |||||||
| Point 1 (pad) | 486 |
|
3.4 | 4.3 | 6.3 | 0.90 | 0.958 |
| Point 2 (centre) | 486 |
|
3.3 | 4.0 | 7.9 | 0.91 | 0.930 |
| Point 3 (fan) | 486 |
|
2.5 | 3.1 | 4.9 | 0.90 | 0.988 |
| Below range | 475 |
|
2.9 | 3.3 | 8.8 | 0.69 | 0.576 |
| Within range | 703 |
|
2.1 | 2.8 | 4.1 | 0.88 | 0.786 |
| Above range | 280 |
|
5.8 | 6.2 | 7.8 | 0.97 | 0.651 |
| Illuminance (lx) | |||||||
| Point 1 (pad) | 486 | 291 | 596 | 804 | 4.3 | 1.01 | 0.996 |
| Point 2 (centre) | 486 |
|
1 663 | 2 733 | 7.4 | 0.96 | 0.980 |
| Point 3 (fan) | 486 | 664 | 664 | 874 | 3.7 | 1.01 | 0.999 |
| Below range | 18 | 145 | 214 | 245 | 6.4 | 0.88 | 0.819 |
| Within range | 1089 | 254 | 713 | 913 | 5.5 | 1.02 | 0.990 |
| Above range | 351 |
|
1 824 | 3 115 | 3.7 | 0.98 | 0.946 |
Humidity mirrored this on the other side of the day. The
calibration window, 16:00–18:00, never saw relative humidity above
65 %, whereas the morning windows reached 82 %. Above 65 % RH the
sensor read 5.4 % RH dry on average (95 % CI 3.6–6.8) against
0.8 % RH dry (CI
Illuminance behaved differently. The relative error was smaller above the calibrated range (MAPE 3.7 %, 95 % CI 1.7–6.1) than within it (5.5 %, 4.2–7.1), and the Bland–Altman analysis found no proportional bias in either group (Fig. 4c). Two features of the data account for the difference. First, the absolute error grew far more slowly than the light itself, from a mean of 0.56 klx below 10 klx to 2.8 klx above 60 klx, so the relative error fell with illuminance, from 7.7 % below 10 klx to 3.3–4.0 % above 20 klx. The fall was already complete within the calibrated range (3.3 % at 20–35 klx), so the lower MAPE above the range reflects the light level, not the calibration. Second, the large errors of the extrapolated range came from a single window rather than from the range as a whole. At point 2 at midday on the first day of the 1 min test, in its 12 readings above 64 klx, the sensor read 15.8 %–18.8 % below the reference, whereas in the 17 readings above 60 klx from the other windows at the same point it read 0.7 %–3.9 % above it. The BH1750 is linear over its range and its calibration slope was close to 1 (0.95–1.07), so unlike the temperature sensor it gives no reason to expect the relation to change above the calibrated range. The under-reading of that window was proportional rather than capped, which argues against saturation of the sensor (full count at about 55 klx in its default setting, extendable by shortening the measurement time [17]) and is more consistent with a difference in exposure between sensor and reference: the BH1750 faced upwards under the conical cover, whereas the luxmeter was held beside the node at a height that was not recorded, so the two could be shaded differently by the cover, the structure or passing cloud. Shading was not measured, and no field note of shading or cloud was kept for that window. The same exposure difference can arise in any comparison between a mounted sensor and a hand-held reference.
The illuminance measured by the BH1750 is photometric: it weights light by the spectral sensitivity of the human eye, whereas crop growth responds to photosynthetically active radiation (PAR, 400 nm–700 nm) and the greenhouse energy balance to total solar irradiance. The conversion between them depends on the spectrum, which changes with cloud, sun elevation and the cover. In this system the BH1750 serves as a light signal against which a controller switches shading or ventilation; its readings do not convert to PAR, solar irradiance or the light supply of crops without a spectral calibration.
Across points, the error of the temperature sensor rose from the pad end to the fan end of the house (RMSE 0.65, 1.07, and 1.58 °C at points 1, 2 and 3). At midday the reference at points 2 and 3 read 2.3 and 1.9 °C warmer than at point 1 and 6.7 and 3.5 % RH drier, consistent with the air warming and drying after it leaves the evaporative pad. The rise, 0.10 to 0.23 °C m−1 of distance from the pad, lies within the 0.06 to 0.27 °C m−1 reported along fan-and-pad greenhouses [3, 6, 29]. Excluding the 43 windows whose reference values carry finer resolution than the instrument displays changed the per-point RMSE by at most 0.11 °C and 0.3 % RH and left the ordering of the range classes unchanged: the share of readings within tolerance stayed highest within the calibrated range (80 and 94 %) and lowest above it (39 and 31 %), although below the range it fell to 64 and 80 % for temperature and humidity.
3.3 Transmission Interval
Because each interval was run on different days (Section 2.5), the effect of the interval was isolated by subsampling the 1 min series of each window at 5 and 10 min, which compares the intervals on identical weather (Fig. 5). For temperature and humidity the interval hardly mattered. Over the 18 windows, the hourly mean from a 10 min series differed from the 1 min mean by a median of 0.05 °C and 0.15 % RH, and by at most 0.22 °C and 0.91 % RH, about a fifth of the reference accuracy. The hourly maximum was underestimated by at most 0.7 °C and 1.0 % RH. The same asymmetry, means robust to sparse sampling and extremes not, appears in daily temperature records sampled at decreasing frequency [30], and a subsampling test of 1 min temperature and humidity logs in museum rooms found hourly logging sufficient there [31]; a greenhouse changes faster than a museum, and still the 10 min series lost little.
Two qualifications bound these results. First, the subsampled series reproduces what the firmware used here did at the longer intervals: it transmitted one instantaneous reading every 5 or 10 min. A node that averages its readings between transmissions would track the hourly mean at least as well, because averaging smooths fluctuations rather than sampling them by chance, but it would flatten the peaks further, so the losses of maxima reported here are lower bounds for such a node. Second, the 18 windows per quantity cover three times of day on two days at three points in one greenhouse. Within an hour, illuminance varied by 6 to 71 klx (median 18 klx), so the windows included both steady light and light varying as under broken cloud, and air temperature reached 47.6 °C. They did not include rain or the operation of shading screens, and the weather of the six days was not recorded; cloud that changes faster or more often than in these hours, or temperatures beyond those observed, could produce larger losses at long intervals.
Illuminance was the exception. Under broken cloud, irradiance at a point can exceed its clear-sky value by more than 50 % on sub-minute time scales [22], and much of the cloud-driven variability lies below one minute [32]. A sparse series samples those changes by chance: the 10 min series shifted the hourly mean by a median of 0.36 klx but by up to 10.1 klx, and missed the hourly maximum by up to 19.9 klx. For a system whose purpose is to trigger shading or ventilation on a light or temperature threshold, a 10 min interval was therefore adequate for hourly means of air temperature and humidity in these windows, with maxima underestimated by up to 0.7 °C, but not for illuminance. The observation guidance of the World Meteorological Organization, which concerns averaged values, treats a 1 min average as the most suitable [21]; a node that sends single readings, as here, has no averaging to smooth the light signal. The result also fits the time scale on which greenhouse climate is controlled, of the order of minutes [33].
3.4 Implications for Low-Cost Greenhouse Monitoring
Within the conditions they were calibrated in, the corrected sensors agreed with the reference to within its stated accuracy; above those conditions they did measurably worse. The practical consequence concerns the calibration window. A two-hour calibration in the late afternoon covered 31 °C–39 °C and 41 %–65 % RH, whereas the house spanned 28 °C–48 °C and 19 %–82 % RH in the performance tests. A calibration that spans a full diurnal cycle, or at least includes a midday and an early-morning hour, would cover the conditions in which a controller acts and should remove most of the extrapolation error seen here, provided the fitted form follows the relation between sensor and reference over the wider range, which may need a non-linear term at high temperature; this implication was not tested here. Similar advice comes from a systematic review of low-cost temperature sensors found that the temperature range covered changes the calibration outcome and called for calibration steps across the range of use [10], and a laboratory evaluation of seven low-cost thermohygrometers advised fixing the calibration range to that of the intended study [19]. Most such calibrations, however, are made indoors or under controlled conditions [10], and a field correction trained at one site degraded when applied where conditions differed [14]. The present data put a size on the consequence inside a working greenhouse: agreement within the reference accuracy fell from 83 to 38 % of temperature readings once the air rose above the calibrated range.
The findings come from one greenhouse, three nodes and one pair of sensor models, compared with reference instruments of limited accuracy that are not traceable, on days whose dates, weather and cooling operation were not recorded. They show what happens to a field calibration of this kind when the conditions leave its range; they do not establish that the same errors, or the same intervals, apply to other tropical greenhouses, other cooling regimes or other sensors. Transfer to another house is most plausible where the same mechanism applies: a low-cost node calibrated in a short window and then used across a wider diurnal range of temperature and humidity.
4. Conclusions
Three low-cost nodes (ESP32 with SHT20 and BH1750 sensors) were calibrated against reference instruments in a fan-and-pad cooled greenhouse in the humid tropics and followed through 1458 paired readings per quantity at three times of day. A two-hour late-afternoon calibration reduced temperature and humidity offsets of up to 2.4 °C and 8.5 % RH to at most 0.34 °C and 2.2 % RH in validation, and within the conditions it covered the corrected sensors agreed with the reference to within its stated accuracy: 83 % of temperature readings and 92 % of humidity readings fell inside ± 1 °C and ± 5 % RH. Because the reference instruments were not traceable, this is agreement with them, not established accuracy. Above those conditions the agreement failed, with 38 % of readings within tolerance above 39.3 °C, mostly at midday, and 36 % above 65 % RH in the morning, and the temperature difference grew with temperature only beyond the fitted range, where the sensor–reference relation steepened and the linear correction could not follow. The late-afternoon window, the one that matched the calibration, gave the smallest errors, a temperature RMSE of 0.48 °C and a humidity RMSE of 1.4 % RH. Subsampling showed that, in the 18 hourly windows tested, a 10 min interval changed hourly means of temperature and humidity by at most 0.22 °C and 0.91 % RH but missed hourly peaks of illuminance by up to 19.9 klx; these figures describe the conditions sampled and may not hold under faster-changing weather. For low-cost nodes in this greenhouse, most of the disagreement on which cooling decisions rest arose outside the calibration window. Whether a calibration spanning a full diurnal cycle, with the sensors in ventilated radiation shields and referenced to an aspirated, traceable instrument, removes the extrapolation error, and whether the pattern holds in other greenhouses and with other sensors, remains to be tested.