1. Introduction
Plants are sensitive to minor changes in environmental conditions, such as light, temperature and humidity. For instance, Arabidopsis is sensitive to as little as a 2°C change in temperature (Balasubramanian et al., Reference Balasubramanian, Sureshkumar, Lempe and Weigel2006). Vertical farms are indoor farms in which multiple layers of planting beds are stacked on top of each other, exploiting the vertical space to enable more crops to be grown per square meter. Vertical farming is considered a kind of controlled environmental agriculture (CEA) practice, because it is possible for growers to customise the growing conditions of crops to maximise yield for each species by controlling aspects of the plant’s environment such as their temperature and light exposure (Farhangi et al., Reference Farhangi, Mozafari, Roosta, Shirani and Farhangi2023), see Figure 1a. Minor environmental (Gavhane et al., Reference Gavhane, Hasan, Singh, Kumar, Sahoo and Alam2023; Ke et al., Reference Ke, Yoshida, Hikosaka and Goto2021; Kong & Nemali, Reference Kong and Nemali2023) differences may also impact agriculturally important qualities other than yield, such as aesthetic qualities, nutrient concentration and taste compounds (Gavhane et al., Reference Gavhane, Hasan, Singh, Kumar, Sahoo and Alam2023; Ke et al., Reference Ke, Yoshida, Hikosaka and Goto2021; Kong & Nemali, Reference Kong and Nemali2023). Farmers need to balance these requirements when designing their optimised vertical farm conditions.
To optimise vertical farm conditions, researchers will usually employ a complete randomised experimental design in which they measure crop traits in several different conditions, a strategy that has been successfully deployed in many studies to optimise light, temperature and humidity separately (Carotti et al., Reference Carotti, Graamans, Puksic, Butturini, Meinen, Heuvelink and Stanghellini2020; Most et al., Reference Most, Lefsrud, Gravel and Azad2019; Wang et al., Reference Wang, Lu, Tong and Yang2016). However, these studies do not consider how changing one growth condition may impact another. Other researchers have utilised multifactorial experimental designs, especially full factorial designs, to combinatorically assess the impact of several environmental variables concurrently (An et al., Reference An, Hwang, Chun, Jang, Lee, Wi, Yeo, Yu and Kwack2021; Ciriello et al., Reference Ciriello, Formisano, Rouphael, De Pascale and Kacira2023; Kamenchuk et al., Reference Kamenchuk, Rumiantsev, Dzhatdoeva, Sadykhov and Kochkarov2023). For instance, a 4 x 4 x 3 full factorial design was used to find optimal combinations of root temperature, air temperature and light intensity for lettuce growth (Carotti et al., Reference Carotti, Graamans, Puksic, Butturini, Meinen, Heuvelink and Stanghellini2020). In contrast, other groups wished to optimise across six environmental parameters at three levels, which would have required 729 experimental treatments using a full factorial design, but they were able to select 27 combinations of conditions to test by using the Taguchi Method to generate an Orthogonal Array of treatment combinations (Farhangi et al., Reference Farhangi, Mozafari, Roosta, Shirani and Farhangi2023). The Taguchi Method is a strategy of selecting a smaller number of combinations of conditions to experimentally test: instead of testing all possible combinations of all the parameters, the Taguchi Method ensures that all possible combinations of every pair of parameters are tested at least once. However, even the Taguchi method requires a relatively large number of treatment conditions, which would be difficult to implement in a small vertical farm. Moreover, any randomised experimental design will assume that treatments are homogenous within each treatment block, when in fact light, temperature and humidity may vary within the physical space due to the layout of lights (Figure 1) and the air flow, among other factors, microclimatic conditions that have been modelled in digital twins of vertical farms (Agati et al., Reference Agati, Franchetti, Rispoli and Venturini2024; González et al., Reference González, Sanchez-Londoño and Barbieri2022).
Figure 1. Photographs of the farm. (a) Arrangement of lights above each bay. (b) Close up of the LED lights. (c) Subdivisions of trays for experiment.
Although the physiological impact of heterogeneous microclimates in vertical farms has not been fully investigated, it has been widely established that in traditional field-based agriculture, conditions are not uniform within a field, resulting in heterogeneity in crops. Soil topography can vary across and within fields, which has been shown to impact how soil can accumulate water and by proxy impact yield (Maestrini & Basso, Reference Maestrini and Basso2018). It has also been shown that increasing the distance from the edge of a field that a crop is grown in can increase yield. Moreover, crop yields can also be impacted by the landscape that surrounds a field and the amount of shading that a crop receives (Fincham et al., Reference Fincham, Redhead, Woodcock and Pywell2023). Within the field variation of the environment is detrimental to both food security and to profits made by farmers. This is because within-field variation impacts how a crop is shaped, its size, colour and yield, all factors that can lead to crops failing quality checks and being disposed of (Ishangulyyev et al., Reference Ishangulyyev, Kim and Lee2019). Vertical farming removes sources of heterogeneity such as soil topography and landscape. However, there are still sources of heterogeneity that occur within vertical farms as shown here.
In this paper, we have taken advantage of the heterogeneity of light intensity and quality that exists in a vertical farm to suggest an innovative observational experimental design for condition optimisation. We measured the position-specific light intensity and quality across 256 different positions in a vertical farm and modelled how these factors impact biomass of different kinds of microgreens. We suggest that exploiting existing heterogeneity in vertical farms to perform observational studies will enable us to optimise treatments using less space-intensive experimental set-ups than in traditional environmental optimization approaches, allowing researchers to test hundreds of different combinations of conditions in a single experiment.
2. Materials and methods
2.1. Measuring light heterogeneity in a vertical farm
This work was performed in the Grow It York vertical farm (Doherty et al., Reference Doherty, Bryant, Denby, Fazey, Bridle, Hawkes, Cain, Banwart, Collins, Pickett, Allen, Ball, Gardner, Carmen, Sinclair, Kluczkovski, Ehgartner, Morris, James and Connolly2022) which uses LettUs Grow aeroponic technology (Chittibomma et al., Reference Chittibomma, Yadav and Reddy2023) in central York. Plants were grown in 0.25 m2 trays, with four trays per bed, with four beds stacked vertically, totalling 16 trays. Above each bed was a set of three LED lighting fixtures, Horti-blade BRWFR-4 spectrum, from Vertically Urban. Due to space constraints during retrofitting, the LED fixtures were positioned unequally above the bed with two fixtures towards the front and one towards the back.
All seed varieties and their sources are found in Table 1. To establish the extent of light heterogeneity within the vertical farm, we divided each tray into 16 unique sections each with an area of 83.7 cm2. Microgreens were also grown in a ring which measured 9cm at the left and right sides of the mat and 11.5 cm at the top and bottom. These rings surrounded the 16 squares in the centre of the mat. The plants in these rings were not measured as part of this experiment to negate edge effects in our samples (see Supplementary Figure S1). Each position was given a unique x and y coordinate and in these same positions, plants would be grown and have their biomasses recorded. Across all beds, this resulted in 256 unique positions. Both photosynthetically active radiation (PAR) and spectral irradiance were measured once within a month of the experiment taking place in these positions at the level of the tray in the absence of plants. PAR was measured using a PAR Special from Skye Industries. Spectral irradiance was measured using an Ocean Fire Spectrometer from Ocean View.
Table 1 Description of microgreen varieties and harvest times
2.2. Finding average blue, red and far-red light measurements
Plants were grown under a spectrum consisting of 22%B 14%G 64%R 7%FR, characterised by blue (400–500 nm), green (500–600 nm), red (600–700 nm) and far-red wavelengths (700–780 nm). For our experimental purposes, we measured narrower waveband ranges: 445–456 nm for blue, 653–668 nm for red and 725–735 nm for far red (Banerjee et al., Reference Banerjee, Schleicher, Meier, Viana, Pokorny, Ahmad, Bittl and Batschauer2007; Butler et al., Reference Butler, Hendricks and Siegelman1964). These experimental wavebands correspond to plant photobiological sensitivity, and the LED emission peaks of the vertically urban lighting system (see Supplementary Figure S2). The average intensities were calculated for each experimental waveband at each of the 256 positions.
2.3. Observational experiment
All plants were grown under 12 hours of light per day which delivered 54.75–62.32 μW/cm2/nm of red light, 37.99–79.36 μW/cm2/nm of blue light and 5.43–14.43 μW/cm2/nm of far-red light. This corresponds to a PAR reading in the range of 263–568 μmol m−2 s−1 for 12 hours a day. These light values are expressed as ranges due to the light heterogeneity in the facility (see Supplementary Table S1). PAR readings are in different units of measurement as spectral measurements, because they are values that are calculated based on the absorption spectra of plants and estimate the quantity of photons that are estimated to be absorbed by the plants per surface area each second. It is challenging to convert to μW/cm2/nm, because conversion requires knowing the distribution of the wavelength of lights. On the other hand, vertical farming light engineers prefer the unit μW/cm2/nm, because it directly relates to energy consumption. For this reason, we chose to maintain different sets of units for PAR and light quality measurements throughout.
Microgreens were grown on Growfelt wool carpet matting. An aeroponic system was used where the pH was 5.9 and the electrical conductivity was 1.7. The nutrient solutions used were Hydromax Grow A and B. The aeroponic system was activated for 1 minute every 5 minutes. Each microgreen variety, as outlined in Table 1, was grown in one tray per bed in the arrangement shown in Supplementary Figure S3. Please note that there are two varieties of kale, distinguished herein by their providers, Kale (CN) provided by CN Seeds and Kale (Tozer) provided by Tozer Seeds. This equalled four trays per variety giving 64 unique measurement positions per microgreen variety. All samples were grown for the period specified in Table 1 which varied by species. Each unique position was harvested individually and biomass was recorded.
2.4. Generating mixed effects models
Our aim was to predict the biomass of microgreens, as this is the yield metric. We used a mixed effects model to predict biomass per position (g/83.7 cm2), using the lme4 package in R (T. Wang et al., Reference Wang, Graves, Rosseel and Merkle2022). We assumed that bed and position within a tray would both be confounding factors influencing biomass. We also assumed that the impact of bed and position within a tray would depend on the species/variety of microgreens. Our baseline model was as follows, where m is the microgreen species/variety (see Table 1), b is the bed and c is the position within the tray (i.e. centre, edge or corner, see Supplementary Figure S1).
Equation 1:
$$\begin{align*}\text{Biomass}_m &\sim \beta_0 + \beta_m + \beta_{b,m} + \beta_{c,m} + e_{b,c,m}\end{align*}$$
Next, we added light intensity (PAR, μmol m−2 s−1) to the baseline model, allowing for light intensity to have a different effect on each microgreen variety.
Equation 2:
$$\begin{align*}\text{Biomass}_m &\sim \beta_0 + \beta_m + \beta_1 \text{PAR} + \beta_{2,m}\text{PAR} + \beta_{b,m} + \beta_{c,m} + e_{b,c,m}\end{align*}$$
Finally, we added light quality (μW/cm2/nm.) to the baseline model, again assuming that each microgreen variety may have different sensitivity to light quality, where B, R and FR represent the mean light intensity in the blue, red and far red light quality bands (μW/cm2/nm), as specified in the previous section.
Equation 3:
$$\begin{align*}\text{Biomass}_m &\sim \beta_0 + \beta_m + \beta_1 \text{B} + \beta_2 \text{R} + \beta_3 \text{FR} \\& \quad + \beta_{4,m}\text{B} + \beta_{5,m}\text{R} + \beta_{6,m}\text{FR} + \beta_{b,m} + \beta_{c,m} + e_{b,c,m}\end{align*}$$
As many plant light sensors detect ratios of light quality, we also developed a model that incorporates light quality ratios, where R:B, R:FR and FR:B represent the log10 ratios red (R), blue (B) and far red (FR).
Equation 4:
$$\begin{align*}\text{Biomass}_m &\sim \beta_0 + \beta_m + \beta_1 \text{R:B} + \beta_2 \text{R:FR} + \beta_3 \text{FR:B} \\& \quad + \beta_{4,m}\text{R:B} + \beta_{5,m}\text{R:FR} + \beta_{6,m}\text{FR:B} + \beta_{b,m} + \beta_{c,m} + e_{b,c,m}\end{align*}$$
To compare the performance of these models we used an approach published by (Zuur et al., Reference Zuur, Ieno, Walker, Saveliev and Smith2009). The models with random effects were compared to the baseline model through separate ANOVAs, with all p-values compared to a 0.05 threshold after a Holm’s correction for multiple hypothesis testing. The Bayesian Information Criterion and Akaike Information Criterion were also assessed.
2.5. Regularising general linear modelling
Equations 2–4 contain a large number of parameters compared to the number of observations, which introduces a risk of overfitting. We sought to use a well-established procedure to select a smaller set of parameters to include in our model. Specifically, a lasso regularisation procedure was employed to select a smaller subset of variables and prevent overfitting, using the glmnet package in R (Engebretsen & Bohlin, Reference Engebretsen and Bohlin2019). Data were standardised prior to fitting as a z-score (each variable was subtracted by its mean in the training data and divided by the standard deviation of the training data), but the coefficients reported here are adjusted to be in the original scale. The initial model specification took the form:
Equation 5:
$$\begin{align*}\text{Biomass}_m &\sim \beta_0 + \beta_m + \beta_1 \text{PAR} + \beta{2,m}\text{PAR} \\& \quad + \beta_1 \text{B} + \beta_2 \text{R} + \beta_3 \text{FR} \\& \quad + \beta_{4,m}\text{B} + \beta_{5,m}\text{R} + \beta_{6,m}\text{FR} \\& \quad + \beta_7 \text{R:B} + \beta_8 \text{R:FR} + \beta_9 \text{FR:B} \\& \quad + \beta_{10,m}\text{R:B} + \beta_{11,m}\text{R:FR} + \beta_{12,m}\text{FR:B} \\& \quad + \beta_{b,m} + \beta_{c,m} + e_{b,c,m}\end{align*}$$
This model was evaluated using a two-level leave-one-out cross-validation approach. For each of the 256 observations, a separate glmnet model was fit after leaving out one observation (i.e. 255 observations used to train the model). To select the lambda parameter of this model that would minimise the mean squared error, a further leave-one-out cross-validation strategy was deployed (using 254 observations for training the models for selecting the lambda parameter, by leaving out one of the 255 observations in the training set).
The output of this procedure was a set of 256 different models, each trained on 255 observations. For each model, we then predicted the biomass of the observation that was not used in training the model. This model-fitting approach allows us to evaluate model performance in a way that is not affected by overfitting by testing our models on observations that were not used to train the model. Moreover, this approach provides us with a collection of 256 fitted values per coefficient, which enables us to evaluate how reliant our coefficient predictions are on the inclusion of any individual data point.
3. Results
3.1. Characterising the heterogeneity of lights in a vertical farm
Our first aim was to quantify the level of light heterogeneity at plant level within the vertical farm. To investigate the extent of this heterogeneity we measured both PAR (Figure 2a) and spectral irradiance (Figure 2b) in 256 unique growing positions.
Figure 2. Heterogeneity of light quantity and quality within the vertical farm: (a) PAR readings and (b) average intensity within the following wavelength bands blue: 445–456, red: 653–668 and far-red: 725–735 are shown across the farm. The beds are vertically stacked and each contains four trays with 16 positions per tray. Each square represents an area of size 83.7 cm2. Note that the plants grown along the edge of each tray are not included in this study and the light was not measured in these positions. (c) The relationship between blue and far-red irradiance and PAR readings, where each point represents a position/square in the images in (a) and (b).
The light intensity (PAR reading) was highest towards the front and centre of each bed (Figure 2a). A similar pattern of light intensity was observed for blue and far red wavebands(Figure 2b), except for the rightmost trays in each bed. For red light, there was greater variation between beds than within beds, with the top two beds having higher levels than the bottom two (Figure 2b). Far-red and blue light intensities were highly correlated at low light settings, but blue light levels saturated at higher light intensities (Figure 2c). Although there was a significant correlation between light intensity and blue and far red light quality, (Pearson correlation test, p<0.001 for both), the Pearson correlation coefficient was low (R = 0.22 and R = 0.21, respectively). These results demonstrate the variation in light quantity and light quality in the vertical farm.
3.2. Light quality is a predictor of biomass
In order to determine whether light quantity, light quality or ratios of light quality were predictors of biomass (the yield metric for microgreens), mixed effect models were constructed and compared to a baseline model that did not include any light-related parameters. Statistics for these comparisons are available in Table 2. The best-performing model was the light quality-based model (Equation 3), performing significantly better than the baseline model, even after Holm’s correction for multiple hypothesis testing (p < 0.05). It also minimised the Akaike Information Criterion (AIC); however, this model performed poorly under the Bayesian Information Criterion (BIC) as BIC places a greater penalty on the number of parameters in the model. These results suggest that light quality can be used to predict biomass and that this is a better predictor of biomass than PAR readings alone.
Table 2 Summary of mixed effect model outcomes
This analysis was unable to determine whether combining information about light intensity and light quality ratios could further improve the model. In the next section, we will try to minimise the complexity of the model but also include combined information about light quality and quantity.
3.3. Variable selection highlights which aspects of light intensity and quality are most predictive of biomass
In order to determine which combination of explanatory variables was most predictive of biomass, we deployed a variable selection procedure, as described in the methods (via lasso regularisation). Our reduced models were able to accurately predict the biomass of samples that were not used in training the model (see Figure 2a), with Pearson’s R = 0.864 and P < 0.05 (P = 7.6e−78). Our predictions were also significantly correlated with the true biomass (p < 0.05) for each of the individual varieties after Holm’s correction, with Pearson’s R of 0.247, 0.487, 0.597 and 0.595 for kale (CN), kale (Tozer), radish and sunflower, respectively. Notably, the variety for which our model performed the worst was kale (CN), which was the lowest biomass variety and also the one with the least variance in biomass.
In our further analysis of model coefficients, we decided to include all explanatory variables that were selected in at least 80% of the models (see Supplementary Figure S4 and Supplementary Table S2). Some light-related variables were equally associated with biomass across all microgreen varieties: light intensity (via PAR reading) and red and blue light quality. Of these, red light quality was strongly negatively associated with biomass in all models, while blue light was positively associated (Figure 3b). Overall light intensity was slightly negatively correlated with biomass in all models (Figure 3b).
Figure 3. Analysis of lasso model coefficients. (a) Here, we compare the actual biomass at each position with the predicted biomass when the lasso model was trained using all the positions except that one (leave-one-out cross validation, LOOCV). Biomass always refers to the total biomass in each 83.7 cm2 sampling site. (b) The histogram of coefficient values for the LOOCV lasso models, for three different parameters (PAR reading, blue and red). (c) The variety-specific coefficients for the R:FR ratio across the LOOCV lasso models. (d) The variety-specific coefficients for kale (Tozer) and radish, when different combinations of beds are used to train the model. Bottom refers to the bottom two beds, while top refers to the top two beds.
Although on their own, light quality ratios did not improve our ability to predict biomass compared to the baseline model, we found several light quality ratio variables that were selected in the combined model. However, these were all specific to individual varieties of microgreens. For instance, separate red:far-red ratio coefficients were selected for each microgreen, with negative associations for both kale varieties and positive associations for radish and sunflower (Figure 3c). A full table of predicted coefficients for all models is available in Supplementary Table S2.
3.4. Variable prioritisation suggests R:FR ratio is a target area in kale and radish production
To further prioritise aspects of light quality impacting yield heterogeneity, we decided to identify light features which were found to be associated with biomass independently in the top two beds and the bottom two beds. The motivation for this is that the top two beds and bottom two beds were found to contain large differences in red light quality (Figure 2b), and we wanted to find variables that influenced biomass under these different red light conditions. To do this, we repeated our two-level lasso regularisation, using only the top two beds and the bottom two beds. The total number of variables selected was reduced, likely because we halved the number of observations used to train the model. However, the red-to-far red light ratio was selected as a key variable for radish and kale (T) in the top two beds, the bottom two beds and in the combined model, with a higher R:FR ratio associated with higher yields in radish and lower yields in kale (T) (Figure 3d). These results confirmed that the R:FR ratio has an opposite direction of association with radish and kale (T) and that this pattern is consistent across the two different red light treatments that these varieties were exposed to. This suggests that the R:FR ratio is an appealing target for further spectral refinement, but that this needs to be performed in a variety-by-variety manner.
4. Discussion
While there are some large vertical farming operations (Stein, Reference Stein2021), many vertical farms (especially R&D vertical farms) are relatively compact, being situated in shipping containers (Schmidt Rivera et al., Reference Schmidt Rivera, Rodgers, Odanye, Jalil-Vega and Farmer2023), in distribution centres (Al-Kodmany, Reference Al-Kodmany2018) or even in retail centres (Martin et al., Reference Martin, Elnour and Siñol2023). For this reason, it may be impractical for these farms to iteratively perform light optimisation experiments, especially factorial designs that investigate the interactions between variables. Moreover, the knowledge obtained from optimising one vertical farm may not translate into optimal conditions in a different setting, as the light-dependency on yield may be dependent on the specific vegetable variety being grown (Cammarisano & Körner, Reference Cammarisano and Körner2022) and other extrinsic conditions, like the temperature, humidity and growth medium (Carotti et al., Reference Carotti, Graamans, Puksic, Butturini, Meinen, Heuvelink and Stanghellini2020). For instance, it has been found that red light can increase or decrease yield, depending on the wider growth context (Wong et al., Reference Wong, Teo, Shen and Yu2020). It has been estimated that up to 85% of the carbon footprint of vertical farms comes from their high electricity demands (Butturini & Marcelis, Reference Butturini, Marcelis, Kozai, Niu and Takagaki2020), so it is paramount that vertical farms find the correct light recipes that effectively balance yield requirements and their carbon footprints.
For this reason, there is an immediate need to develop experimental designs that would prioritise the light qualities that are worth further investigation. We propose a three-phase workflow: (i) quantification of light heterogeneity within the facility (ii) an observational experimental design (iii) variable selection using lasso regularisation. The variables that are consistently selected and have high coefficients could be targeted for further investigation. Critically, we use cross-validation to ensure that our models are able to predict biomass in data that was not used to train the model, which ensures that our results are not simply a consequence of over-fitting. Moreover, this workflow can suggest areas where heterogeneity in microenvironments could be most impacting the yield of the crops, areas warranting technological improvements to ensure greater environmental refinement.
Our work highlights several key associations among light intensity, light quality and biomass. First, heterogeneity in light quality within a vertical farm may have a greater impact on uniform crop yields than heterogeneity in light intensity. This is consistent with (He et al., Reference He, Qin and Chow2019), which finds that light quality influences leafy green properties, even under uniform light intensity conditions. This suggests that it is important for small-scale vertical farms to track the heterogeneity of light quality in their facilities, instead of solely relying on PAR readings. Additionally, sensitivity to light ratios is likely to be variety-specific, which is consistent with several studies that have found different optimal light quality ratios were optimal for different varieties (Jin et al., Reference Jin, Li and Priva2023; Pennisi et al., Reference Pennisi, Orsini, Blasioli, Cellini, Crepaldi, Braschi, Spinelli, Nicola, Fernandez, Stanghellini, Gianquinto and Marcelis2019). This highlights how important it is to have ways of quickly screening light sensitivities in small-scale vertical farms, as a light recipe that works well for one variety is not guaranteed to produce optimal results for another. Other data sets in vertical farms have explored the role of light quantity and quality on yield, but these experiments were performed in a randomised controlled experimental design and so fewer combinations of species and light combinations were tested, with many examples of vertical farm experiments only testing 3–5 different light recipes (Carotti et al., Reference Carotti, Pistillo, Zauli, Pennisi, Martin, Gianquinto and Orsini2024; Dou et al., Reference Dou, Niu, Gu and Masabni2020; Semenova et al., Reference Semenova, Smirnov, Dorokhov, Proshkin, Ivanitskikh, Chilingaryan, Dorokhov, Yanykin, Gudkov and Izmailov2022)
The specific light sensitivities of the microgreens that we highlight are likely to be specific to the vertical farm facility and microgreen varieties we have grown. Nevertheless, they indicate the kind of lessons that could be learnt from performing an observational study within a small-scale vertical farm. Our results have helped suggest light treatments that we can now test on our vertical farm to improve yield and reduce energy consumption. An important caveat is that temperature and humidity heterogeneity were not quantified in our experiment. However, our experimental design could be extended to include these variables.
Open peer review
To view the open peer review materials for this article, please visit http://doi.org/10.1017/qpb.2025.10003.
Supplementary material
The supplementary material for this article can be found at http://doi.org/10.1017/qpb.2025.10003.
Data availability statement
All data and code to produce all figures and statistical tests in the manuscript are available at https://github.com/stressedplants/VerticalFarmOptimisation.
Acknowledgements
We would like to acknowledge A. L. Tozer Limited for providing us with the kale seed. Additionally we would like to thank Paul Scott, Jason Daff, Harry Stevens, Alison Fenwick, Jacob Woodward, Dave Grimshaw and Peter Smithson of the University of York Department of Biology horticulture department for their assistance and support with this work. The research presented in this paper was conducted while PS was employed by Vertically Urban; PS is currently affiliated with The UK Agri-Tech Centre. The authors would like to thank Vertically Urban for their support during the research period. We also thank Spark:York CIC for their support of the Grow It York farm.
Author contributions
W.C. and D.E. conceived the study. W.C., D.E., P.S. and K.D. designed the study. W.C., A.K. and A.T. performed the experiments. W.C., E.R., G.V. and D.E. performed data analysis. W.C. and D.E. wrote and compiled the manuscript with contributions to the text from P.S., K.D., A.K., A.T., G.V. and E.R.
Funding statement
We would like to acknowledge the following funding sources: the Royal Society (RGS/R2/212345: D.E.), Biotechnology and Biological Sciences Research Council (Responsive Mode) (BB/V006665/1: D.E. and W.C.), the Biotechnology and Biological Sciences Research Council (White Rose Doctoral Training Partnership) (BB/T007222/1: E.J.R. and G.Y.W.V.), GenerationResearch and the Department of Biology (W.C., MRes studentship, https://generationresearch.ac.uk/) and FixOurFood - UKRI Strategic Priority Fund Transforming UK Food Systems project FixOurFood (BB/V004581/1: K.D.).
Competing interests
A.L. Tozer Limited supplied the seeds and Vertically Urban Limited provided the lighting.
Open access
Comments
Dear Editor,
Please consider our invited manuscript entitled ‘Harnessing light heterogeneity to optimise controlled environment agriculture’. Here, we suggest an observational study design to prioritise variables for optimising light regimes in vertical farms. We introduce a workflow that relies of regularised regression techniques. This work fits within the scope of the journal because it applies quantitative methods in an innovate way to an area of plant biology. We believe that this work could have wide impact in vertical farms, because many of these farms are small and do not have R&D facilities, but our approach enables them to optimise their light regimes in a streamlined way. Please note that we have a few conflict of interests to disclose: Tozer donated the seeds to us and Vertically Urban provided us with the lights.
Thank you for considering our manuscript.
Sincerely,
Dr. Daphne Ezer
Lecturer in Computational Biology
University of York