Whitepaper

Next level Data-Driven Growing

Taking the next step in data analysis for crop growth optimization
2 | www.letsgrow.com
Introduction
For a grower, it is not unthinkable that his crop management system becomes his best full-time crop consultant
in the nearby future. Think of a crop monitoring system that sends sensor data about the water, energy, and
assimilates balance of the plants to a central data platform.

Based on all this data, the system takes decisions and This takes time and commitment but is indispensable for
controls the greenhouse climate in such a way that crop a correct analysis. Only with a high- quality dataset it is
growth is optimized. This may sound like science fiction, possible to complete the last complicated steps in data
but it is becoming a reality. The first steps have already analytics. Whereby algorithms and models play a decisive
been taken and are described in the previous whitepaper role.
about “Data-Driven Growing in practice.” The focus was
on the following questions: ‘What happened?’ and ‘Why
did it happen?’.

This whitepaper describes the next steps: ‘What will

happen?’ and ‘What is the best that could happen?’.
Knowing the best thing that could happen is the ultimate
goal for crop growth optimization.

As already stressed in the first whitepaper about “Data-

Driven Growing in practice,” data quality is crucial.
Creating a high-quality dataset requires regular calibration
of sensors, consistent crop registrations, and also data
cleaning. Different levels in data analytics

Whitepaper | 3
Algorithms and models
The words “algorithm” and “model” are often used in the context of Artificial Intelligence (AI).
What do these buzzwords mean precisely?

Data Algorithm Model

An algorithm is a method for solving a problem, this method can include formulas.

We all know what a formula is: it calculates an output In this sense, this model still simulates a static, entirely
value from a specific set of variable inputs using several predictable, and well-understood process.
fixed coefficients that describe the exact relationship
between the inputs and the output. When the number of variable inputs increases and the
relationships between inputs and output becomes less
Therefore, such formulas can be used for physical and static and maybe even timedependent and non-linear, and
chemical processes that behave quite steady and where the characteristics of the system vary over time, formulas
the number of inputs is restricted. and static models are no longer usable to simulate the
behaviour of a process or system. That is where AI
In case the description of our process consists of techniques come in.
multiple formulas, we can speak of a model.

4 | www.letsgrow.com
Applying AI in horticulture
Growth and development of a living crop in a greenhouse is an excellent example of a complex and dynamic
system that requires a different approach. The behaviour of a crop is not only determined by the actual
conditions but also depends on many other factors resulting, amongst others, from the crop’s history

To describe these complex and the network is capable of simulating strategy? And what if I change this
dynamic systems, we need the help this behavior when it is fed specific parameter?
of dynamic algorithms and models, with new input data. We can also
based on AI techniques. In general, say: the trained network has become And what is likely to happen this
artificial intelligence tries to mimic a model of the real crop. upcoming week based on the
human intelligence. Looking at plant production systems current weather forecast? A yield
we can think of models that prediction model, for example, can
One of the crucial elements is that describe growth and development, provide useful information to the
it can learn from history by using models that predict yield and quality, grower on the number of kilos
neural networks. These networks models that calculate expected that he is likely to produce in the
can be trained by large sets of energy- and water demand, and next couple of weeks. This is very
recorded input- and output data models that estimate the risks important to conclude profitable
from the crop under research. This of pests and diseases. Different contracts with costumers. An energy
is also called “machine learning,” models together form a decision demand model can help to do
or sometimes “deep learning.” support system that is capable of accurate planning for the purchase
During the training process, the answering the questions that of gas and electricity, also based
network gradually learns how the growers like to ask, such as: what on the weather forecast.
real crop behaves, and eventually, will happen if I maintain my current

Artificial Intelligence Machine learning Deep learning

Any technique wich enables Subset of AI techniques which Subset of ML which make the
computers to mimic human use statistical methods to computation of multi-layer
behavior. enable machines to improve neural networks feasible.
with experiences.

Different terms in the field of artificial intelligence explained.

Whitepaper | 5
Some further considerations regarding AI and models
Models are a simplified representation of a complex reality. Therefore, they can never be perfect for various
reasons. One is that, most likely, our input data is not complete; there might be more factors that influence the
crop then we measure or even are aware of. Furthermore, the behaviour of the real crop can keep on changing
over time, while the model has become static from the moment that we stop training it.

Fortunately, we can keep on teaching the model with new different factors into account, both current and historical,
data when this becomes available. to derive a specific outcome, that we, as human beings,
fail to understand what exactly happens inside fully. Why
Nevertheless, the knowledge inside the model is always does the yield prediction equal 2.3 kg/m² for next week?
based on history, and thus it may fail to predict the future.
Models are developed for a specific purpose within Why does the model expect an increasing botrytis risk?
a restricted range of conditions, so, a model must be We cannot track this exactly. This is in huge contrast with
appropriately used respecting its limitations.Models are old fashioned descriptive/explanatory models build of
never better than the data/information that was used formulas, where every relationship between the inputs
to develop and train them. The general rule of “garbage and final output is fully traceable.
in = garbage out” fully applies to any model, regardless
of how intelligent or advanced it may be. However, Therefore, the use of AI models always requires
this only works in combination with professional data excellent domain knowledge of the application itself. The
management, fast databases, and massive computing outcomes of the model must be assessed and checked
power. The larger the dataset, the more complex the by human experts as well, to detect and filter out non-
model, the more expertise and computing power is realistic results. Only when these modern techniques are
needed. One of the downsides of the AI models is their combined with knowledge about plant physiology and
“black-box” character. The model can take so many physics, excellent results can be expected.

Difference between traditional programming of models and machine learning models.

6 | www.letsgrow.com
Examples of AI applications

In a modern commercial greenhouse, multiple sensors can generate data. For instance, aspirator boxes and
wireless sensors measuring Temperature, Humidity, CO2, PAR light. Meteo sensors measure the weather
conditions outside. Additionally, there may be cameras and weighing scales. Also, the current generation of
climate computers offers large amounts of data about actuators such as ventilation windows, screens, irrigation
valves, etcetera.

Plant sensors and crop registration systems provide data brains by just looking at raw data and simple graphs.
about crop development, yield, and quality.
Both LetsGrow.com and Hoogendoorn have been working
For a grower, it becomes more and more complicated and on various projects trying to get the most out of the
time-consuming to combine this huge amount of data available data with the right tools/ models to optimize
to useful information for decision making and process crop growth and greenhouse climate.
optimization.
The main goal is to increase both production and
Therefore, extra tools are needed. With the assistance of maximum resource use efficiency, hence improve the
smart and advanced data analytics techniques, it is now Return of Investment (ROI). In the next chapter, some
possible to make analysis that cannot be made by human successful projects are discussed.

Whitepaper | 7
Crop growth analysis (Light Use Efficiency)
As a grower or investor, you may observe significant production differences between different greenhouse
locations. Most likely, you want to know the reason behind this. Or suddenly a particular pest or disease appears
during the growing season affecting crop growth and productions. And you want to know: why now? What has
been different this year compared to the previous growing season?

Extensive data analysis of actions and results from As is widely known, photosynthesis plays the key role
the past years can provide answers and insights to the here; the more light can be converted to assimilates by
questions mentioned above. These answers will help to photosynthesis, the more assimilates available for growth
keep on optimizing crop growth. The first question always and production. Photosynthesis is the process that
is: What is your relative cultivation performance? The converts CO2 and water H2O into assimilates (sugars)
Light Use Efficiency (LUE) calculation was developed with the help of energy from PAR light that is intercepted
to answer this question. The LUE shows how efficiently by the leaves.
available radiation is converted into products such as
fruits or flowers.

An example of the light use efficiency during a growing season for two greenhouse lacations

To optimize photosynthesis, multiple factors need to be That is where the LUE model can give clues by calculating
aligned and optimized, both below and above the soil. For the realized LUE and comparing this with the assimilates
example, disturbance of the energy and water balance of balance-related climate factors such as radiation,
the plant cause closing of the stomata, which hampers temperature, RH and CO2 but also with plant data such as
the uptake of CO2 and consequently decreases LUE. fruit set rate and plant load.
For that reason, also controlling irrigation, greenhouse
temperature, and relative humidity (RH) are crucial for Thanks to advanced data analysis tools, it is feasible to
stomata opening. The art of optimizing photosynthesis detect the bottlenecks in the cultivation process and to
is, therefore, detecting and improving the “limiting provide useful insights to the grower on what measures
factor,” which is quite challenging because of the many should be taken to obtain better results.
influencing factors that are continually varying.
Using the LUE analysis, we also provide more insights and

8 | www.letsgrow.com
guidelines to the grower regarding the optimal utilization The scope of this analysis depends on the research
of his equipment. For example, recommendations on questions. Currently, several studies have been conducted
optimal screen usage or window position (ventilation) can for growers worldwide. One data analysis showed that
be given. By the way, the climate monitor (explained in the the LUE fluctuates too much during a cultivation cycle.
previous whitepaper) is a great tool to assist the grower Consequently, this grower decided to first start with
with implementing these recommendations. You can extensive training on Data-Driven Growing and Plant
observe daily (even on a 5-minute basis) if your climate Empowerment, to better master integrated crop and
factors are in balance. climate management. Another data analysis clarified the
differences in production between
Also, crop management can be involved. For instance, multiple greenhouses.
how to plan the plant load relative to yearly radiation?
Maybe proactive fruit thinning is advised during the A third analysis revealed the cause of the unexpected
autumn period. Moreover, how about irrigation? Are EC occurrence Fusarium. These are
levels within a proper range? Most of the new insights a few examples that show how advanced data analysis
based on this integrated analysis can be directly helps growers to gain new insights regarding their
implemented. specific issues, and provides them with information on
how to further optimize cultivation results.

Whitepaper | 9
Improve the greenhouse climate profile
All modern greenhouses have one or more aspirator boxes per climate department measuring temperature,
humidity, and CO2. This data is needed for controlling the greenhouse climate.

Typically, each aspirator box covers quite a large area, cold spots. The heatmap also shows if the greenhouse
for instance, a quarter or half a hectare (2.500 - 5.000 climate is homogeneous or if significant, for example,
m²). The reading of the temperature, let us say 20.5 °C, temperature differences occur.
is assumed to be valid for this entire area. In reality, the
greenhouse temperature across this area may easily vary For making the greenhouse climate more homogeneous,
between let us say 18°C - 24°C. it is essential to find the cause of disturbances. Therefore,
multiple questions must be answered, such as: What
Especially in large greenhouses, the air circulation can is the effect of the wind direction on the temperature
be quite strong, disrupting a homogeneous greenhouse distribution?
climate. Installing wireless sensors inside the greenhouse
allows accurate measuring of the horizontal and vertical What wind speed causes the most significant
distribution of air temperature, relative humidity, and temperature differences? What happens when the
CO2 concentration. Using these wireless sensors in the windows are opening or closing? What is the influence
greenhouses produces vast amounts of data. Let us of screens? For answering these questions, more data
make a simple calculation. Using 100 sensors sampling is needed from different sources. That is why data from
greenhouse temperature and relative humidity every 5 the wireless sensors, the climate computer, and outdoor
minutes results in more than 50.000 data samples per conditions sensors must be combined.
day. An excellent data pipeline is required to handle this
enormous data flow. And data collection is only the start; Visualizing all this data on one dashboard provides
how to extract useful information from these extensive endless possibilities. Using specific filters enables
datasets? answering the questions above. By filtering out only the
moments that the wind direction was southwest, for
The heatmap is a helpful tool to visualize the data. It example, clarifies the effect of a southwest wind direction
shows the hot and cold areas in the greenhouse directly. on the climate profile. It is also possible to zoom into the
The location of these warm and cold spots are not fixed moments with extreme temperature differences in the
but can move across the greenhouse area. Watching the greenhouses and to search the most dominant factor
heatmap images in the form of a time-lapse gives insight causing this temperature difference.
into the appearance and disappearance of these hot and
By asking the right questions, selecting the right filters,
and by looking critically at the data, new insights pop up
regarding the greenhouse climate profile.

Based on these new insights, the proper actions can

be taken. The type of action depends on the issue
involved and also on the characteristics of the specific
greenhouse. Sometimes a new screen is required, which
is less porous to better insulate against heat loss.
In other cases, only the screen strategy needs to be
adjusted, mostly aiming to minimise the occurrence of
screen gaps. Sometimes, the operation of the existing
fans needs to be optimized.

This heatmap shows that there are significant temperature

differences within the greenhouse.

10 | www.letsgrow.com
In one case, the grower decided to change the ventilation controlled greenhouse climate. This way, the data
of his greenhouse, based on the outcome of the analysis. analysis has resulted in an optimized greenhouse climate
Now the ventilation windows have been divided into profile, which has a substantial positive effect on crop
more controlled groups, thus obtaining a more locally growth, plant health, and production.

A dashboard created to find the cause of a heterogeneous greenhouse climate.

Yield prediction using machine learning

Having an accurate yield prediction is essential in the horticultural industry. A precise yield prediction enables
the grower to negotiate on better conditions and higher product prices and hence, increasing his profits.

In general, with yield prediction, the sales process can Other growers make use of historical data, combined
be optimized. Preventing shortages and surpluses of the with weather forecasts. In most cases, however, due to
product gives better returns on sales. It is also possible the inaccuracy of the prognosis, the amount of expected
to optimize the internal processes at the nursery, such as available product in the coming period is still too
labour scheduling. uncertain. Typically, the forecasts regularly differ a few
dozen percents from reality, making reasoned decisions
Growers, both individual and in growers associations, impossible.
and traders have already been struggling with yield
predictions for many years. Some growers just walk So the question is: how can we make yield prediction as
through the greenhouse and based on visual observations accurate as possible? Several types of yield prediction
of the crop, a yield prediction for next week can be models can be distinguished. One type is the model that
established. However, such a prediction is likely to be far includes every single step of the production mechanism
from consistent and not adequate to be used for serious from radiation input to production output. Therefore,
sales planning. these models consist of multiple sub-models like
photosynthesis, sink-source ratios, and light interception.

Whitepaper | 11
Such a model can be highly accurate but typically needs are filled in by the grower, a fairly good estimation can be
many carefully adjusted parameters to function correctly. made of the yield for the upcoming weeks.
This makes them less suitable for use in daily practice.
However, the research experience behind these models While this model is used by growers for decades and
can be beneficial to develop a more straightforward and provides relatively accurate predictions, it is still not
more practical yield prediction model. accurate enough. Besides, it requires a commitment to
do all (extra) registrations correctly every single week.
The first yield prediction model was developed almost By applying AI techniques such as machine learning, we
20 years ago. This plant physiologic based model makes managed to develop a new yield prediction model that
use of the well-understood relationship between averaged performs with excellent accuracy: better than 90% on
temperature and ripening time. Based on crop registration average.
data and the expected greenhouse temperature, which

Yield predictions with machine learning compared with realized production.

This new approach is based on the large data sets that use of crop registrations which are done manually by
many growers have been building up over the years. human persons. Inaccurate records will directly affect the
These data sets typically contain crop registrations, accuracy of the outcome of the model. The prediction can
climate data, and data regarding production and quality. never be better than the quality of the input.
This is the ideal starting point for machine learning, which
is able to find the correct relationships between crop Therefore, the future of yield prediction is in using new
status, greenhouse climate, and production, depending on techniques to obtain information from the crop, instead of
a specific cultivation strategy and type of plant. manual crop registration. Think about cameras mounted
on moving robots, which can observe, count, and assess
That is why the model must be trained per grower, using all fruits in the greenhouse much more accurate and more
his private data. Consequently, every grower gets his own, frequently than is feasible by human persons. Combining
tailor-made prediction model that provides surprisingly all these data will step by step further increase the
accurate yield predictions for 1-4 weeks ahead. It has accuracy of yield prediction models.
to be stressed, however, that also this model makes

12 | www.letsgrow.com
Stomata conductivity module
In practice, many crops experience stress due to different reasons, such as water shortages or extreme
conditions in terms of high irradiation in combination with high greenhouse temperature and/or low RH.

As stomata are the only physical and mechanical The high RH levels will prevent closing of the stomata.
instruments of plants to control their water balance, the
behaviour of the stomata reveals if, and to what extent The stomatal conductivity module is a so-called smart
the plant is suffering water stress. In the case of water sensor that consists of a combination of real sensors
shortage plants will close their stomata to decrease the (hardware) and a calculation model (software). It
loss of water. While this prevents the plants from drying simulates/calculates stomatal conductance based on the
out, it also reduces photosynthesis because the exchange energy balance and water balance of the leaves of any
of CO2 is hampered. For increasing the production of plant. This is a physical process which can be described
assimilates, it is crucial to create a greenhouse climate by a set of static formulas. AI techniques are not
that is optimal for photosynthesis. required. Since its first introduction in 2007, the model
has proven its value in multiple research projects and
In general, this means that the air temperature, CO2, and practical applications.
RH must be aligned with the intensity of solar radiation.

This graph shows the measurements and derived calculations, amongst other things evaporation rate and stomatal conductance,
as a result of plant water status and climate conditions.

The sensors measure temperature and RH of the air, partially open, while a higher value indicates that stomata
plant leaf temperature, and PAR intensity. The software are more open. With this module, the water status, as a
model now calculates the evaporation rate, based on result of water uptake and evaporation, can be monitored
the energy balance of the crop, expressed in g/m².h. from the plant’s perspective. As a grower, you can see
Combined with the current vapour pressure difference and learn how the stomata respond to the various
(VPD), the stomatal conductivity can be determined and climate factors. This way, it is possible to increase
presented on a graph. The chosen unit is kg/ kPa.m².sec photosynthesis and reduce plant stress by optimizing the
to indicate the ratio between evaporation rate and VPD. A greenhouse climate. The stomata conductivity module is
low conductivity value means that stomata remain only available for different types of crops.

Whitepaper | 13
 Conclusion
There are several ways to add value to data.

Using models is one of the most important tools to do

this. Examples are the stomata model that calculates
stomata conductivity using static formulas, or yield
prediction based on a machine-learning model. The goal
is always to create information from the data that can be
used to optimize plant growth. It all starts with having all
the correct data available at a central place.

Next, data quality needs to be of a sufficient level,

because of the rule garbage in = garbage out. This
applies to any model. Lastly, any model can only be used
if it is easily accessible. This means that the model needs
to be available 24/7 for every client. For all these reasons,
it is vital to use a central data platform. Only by using an
excellent high performance data platform, it is possible to
implement next level data-driven growing in practice fully.

14 | www.letsgrow.com
Epilogue
Hoogendoorn Growth Management creates
sustainable and user-friendly automation solutions
for every kind of horticultural business worldwide.

It provides growers with a complete solution to efficiently

manage the greenhouse climate, irrigation, and energy
usage, regardless of the greenhouse structure, location,
and equipment. The intelligent controls enable growing
high-quality crops and achieve maximum crop yields with
minimum use of scarce resources such as water, energy,
and nutrients. The main focus is on growth, continuity,
and innovation.

LetsGrow.com has been operational since 2002.

Together with Wageningen University Horticultural
Research, models for forecasting crop yields have been
developed. However, the services of LetsGrow.com span
horticulture in its entirety: all popular brands of climate
computers are being supported as a data source. Clients
have 24/7 access to real-time data all around the world.
The fact that already 1000 greenhouse growers have
a subscription to LetsGrow.com is a testament to this.
This includes both individual companies as well as major
growers associations.

Both Hoogendoorn and LetsGrow.com offer the

possibility to collect and to analyze cultivation related
data. Via an online platform, growers can analyse data
regarding e.g., greenhouse climate, crop, labour, and
energy consumptions. Graphs and dashboards can easily
be created to gain insight into the cultivation data. Also,
calculations upon real-time data can be performed to
deliver even more insight.
This way, data is transformed into meaningful
information.

Whitepaper | 15
 Westlandseweg 190
3131HX Vlaardingen
The Netherlands

T +31 10 460 81 08
[email protected]

16 | www.letsgrow.com