Using Deep Learning to Predict Plant Growth and Yield in

Greenhouse Environments
ABSTRACT

Effective plant growth and yield prediction is an essential task for greenhouse growers and
for agriculture in general. Developing models which can effectively model growth and yield
can help growers improve the environmental control for better production, match supply and
market demand and lower costs. Recent developments in Machine Learning (ML) and, in
particular, Deep Learning (DL) can provide powerful new analytical tools. The proposed
study utilises ML and DL techniques to predict yield and plant growth variation across two
different scenarios, tomato yield forecasting and Ficus benjamina stem growth, in controlled
greenhouse environments. We deploy a new deep recurrent neural network (RNN), using the
Long Short-Term Memory (LSTM) neuron model, in the prediction formulations. Both the
former yield, growth and stem diameter values, as well as the microclimate conditions, are
used by the RNN architecture to model the targeted growth parameters. A comparative study
is presented, using ML methods, such as support vector regression and random forest
regression, utilising the mean square error criterion, in order to evaluate the performance
achieved by the different methods.
INTRODUCTION

As with many bio-systems, plant growth is a highly complex and dynamic environmentally
linked system. Therefore, growth and yield modeling is a significant scientific challenge .
Modeling approaches vary in a number of aspects (including, scale of interest, level of
description, integration of environmental stress, etc.). According to (Todorovski and
Dzeroski, 2006; Atanasova et al., 2008) two basic modeling approaches are possible, namely,
"knowledge-driven" or "data-driven" modeling. The knowledge driven approach relies
mainly on existing domain knowledge.

In contrast, a data-driven modeling approach is capable of formulating a model solely from

gathered data without necessarily using domain knowledge. Data driven models (DDM)
include classical Machine Learning techniques, artifical neural networks (Daniel et al.,
2008), support vector machines (Pouteau et al., 2012), and generalized linear models. Those
methods have many desirable characteristics, such as imposing fewer restrictions, or
assumptions, the ability to approximate nonlinear functions, strong predictive abilities, and
the flexibility to adapt to inputs of a multivariate system (Buhmann, 2003). According to
Singh et al., 2016 and reviewed by Liakos et al., 2018 Machine Learning (ML), linear
polarizations, wavelet-based filtering, vegetation indices (NDVI) and regression analysis are
the most popular techniques used for analyzing agricultural data. However and besides the
aforementioned techniques, a new methodology which is recently gaining momentum is deep
learning (DL)(Goodfellow et al., 2016).

DL belongs to the machine learning computational field and is similar to ANN. However, DL
is about “deeper” neural networks that provide a hierarchical representation of the data by
means of various operations. This allows larger learning capabilities, and thus higher
performance and precision. A strong advantage of DL is feature learning, i.e., automatic
feature extraction from raw data, with features from higher levels of the hierarchy being
formed by composition of lower level features (Goodfellow et al., 2016). DL can solve more
complex problems particularly well, because of the more complex related models (Pan and
Yang, 2010). These complex models employed in DL can increase classification accuracy
and reduce error in regression problems, provided there are adequately large data-sets
available describing the problem.

LITERATURE SURVEY

Some of the most valuable documents and papers after extensive background work are:

S.V Baria [1] built up a strategy to demonstrate how a high- resolution satellite imagery is
essential to isolate rice cultivation. LAI’s multi-regression model was taken as an input
and NDVI or any other vegetation index calculated from red and near-infrared spectral
reflection was taken as another influence under normal environmental conditions and
common agronomic practices. The top methodology of estimating rice yield using
satellite imagery could be during the period of maximum vegetative growth.

SnehalS.Darikar et al. [2]discussed in their paper the use of Artificial Neural Network to
predict crop yield. The paper senses the parameters of the regional soil and the various
atmospheric conditions. Then it furthers analyses by using feed forward back
propagation ANN. By using Mat lab ANN approach was made more efficient. They
structure a system that accurately links climate effects to crop yield, can be used to
estimate long - term or short-term crop production and can also obtain an ANN with
adequate and useful data.

AgaarwalSachin [3] discovered the air quality index by utilizing neural system-based
air quality factors, which worked with a fractional number of informational indexes
and are sufficiently strong to deal with information with noise and inaccuracies. Various
accessible varieties of neural system models, for example, Repetitive Network Model
(RNM), Consecutive System Development Demonstrate (SNCM), Change Point Discovery
Display with RNM (CPDM), and Self Sorting out Element Maps (SOFM) are
executed in this paper for anticipating air quality. The created models were utilized
to reproduce and figure the air quality index dependent on long haul (yearly) and in addition
present moment (every day) informational collections. The models can accurately anticipate
air quality patterns.
B.A. Smith et al. [4] discussed about all year or long-haul atmospheric temperature
expectation models that were produced for estimating forecast horizons of 1 to 12 h
utilizing Ward-style Artificial Neural System. These models were intended for general
support in decision making. The variations of the ANN plan described here provide
greater precision compared to previously developed winter models amid the winter time
frame. The models that had included precipitation terms likewise as a part of the air
prediction model in the input vector were progressively exact.

B. J I et al. [5] The aim of this study was to: (1) research whether Artificial neural
system (ANN) models could effectively and efficiently forecast Fujian rice yield for
characteristic mountainous climate and atmospheric conditions, (2) assess the
performance of the ANN model in comparison to varieties of rising parameters and (3).
Compare the effectiveness of multiple linear models of regression with models of ANN. The
models were developed using historical harvest data from several locations in Fujian Field-
explicit rainfall information and the climate conditions were utilized at every location
for the rice yield prediction.

Lillian Kay Peterson [6] Here, Lillian created satellite investigation methodologies and
programming devices to forecast crop yields two to four months ahead of time, before
the harvest. This procedure estimated relative vegetation condition dependent on pixel-
level customary irregularities of NDVI, EVI and NDWI indices. Since no crop mask,
modification, or sub-national ground truth information are vital, this procedure can be
valuable to any area, location, harvest, or climate, making it ideal for African nations
with little fields and poor ground perception.
PROPOSED SYSTEM

In the proposed approach the DL and ML models were applied to Ficus plants
(Ficusbenjammina) data collected from four cultivation tables in a 90𝑚2 greenhouse
compartment of the Ornamental Plant Research Centre (PCS) in Destelbergen, Belgium.
Plant density was approximately 15 pots per 𝑚2 , where every pot contained 3 cuttings.
Greenhouse microclimate was set by controlling the window openings, a thermal screen, an
air heating system, assimilation light and a CO2 adding system. Plants were irrigated with an
automatic flood irrigation system, controlled by time and radiation sum. Set points for
microclimate and irrigation control were similar to the ones used in commercial greenhouses.
The microclimate of the greenhouse was continuously monitored. Photosynthetic active
radiation (PAR) and CO2 concentration were measured with a LI-190 Quantum Sensor (LI-
COR, Lincoln, Nebraska, USA) and a carbon dioxide probe (Vaisala CARBOCAP GMP343,
Vantaa, Finland), respectively. Temperature and relative humidity were measured with a
temperature and relative humidity probe (Campbell Scientific CS215, Logan, UT, USA),
which was installed in a ventilated radiation shield. Stem diameter was continuously
monitored on these plants with a linear variable displacement transducer (LVDT, Solartron,
Bognor Regis, UK) sensor. The hourly variation rate of stem diameter (mm 𝑑 −1 ) was
calculated as the difference between the current stem diameter and the stem diameter
recorded on hour early for a given time point.
MODULES IDENTIFIED

ML module for growth prediction

One of the main advantages of Machine learning (ML) techniques is that they are capable of
autonomously solving large non-linear problems using datasets from multiple sources. ML
enables better decision making and informed action in real-world scenarios without (or with
minimal) human intervention.

In recent years different ML techniques have been implemented to achieve accurate plant
growth, yield and production prediction for different crops. As already mentioned, the most
successful techniques are Artificial Neural Networks, Support Vector Regression (SVR),
M5-prime Regression Trees, Random Forests (RF), and K-Nearest Neighbors (Chlingaryan
et al., 2018). In this project, SVR and RF models as baseline models to predict plant yield
and growth are used.

Support vector regression (SVR)

Support vector regression (SVR) arises from a nonlinear generalization of the Generalized
Portrait algorithm developed by Vapnik (Cortes and Vapnik, 1995). It projects the input data
into a higher dimensional space using a kernel function and separates different classes of data
using a hyperplane. The trade-off between margin and errors is controlled by the
regularization parameter c. SVR with radial basis kernel functions(SVRrbf)uses 𝐾(𝑥𝑖 ,x𝑗) =
exp (−𝑦‖𝑥𝑖 − 𝑥𝑗‖ 2 ). Here y is a constant used in the radial basis function.

Random forest (RF)

RF belongs to the category of ensemble learning algorithms, having been proposed by Ho in

(Ho, 1998). As a base learner of the ensemble, RF uses decision trees. The idea of ensemble
learning is that a single predictor is not sufficient for predicting the desired value of test data.
The reason being that, based on sample data, a single predictor is not able to

Deep Learning extends classical ML by adding more "depth" (complexity) into the model, as
well as transforming the data using various functions that create data representations in a
hierarchical way, through several levels of abstraction. A strong advantage of DL is feature
learning, i.e., automatic feature extraction from raw data, with features in higher levels of the
hierarchy being formed through composition of lower level features

The implemented approach involved three steps:

1. Data preprocessing and data cleaning.

2. Data splitting into training, validation and test datasets.
3. DL/LSTM, SVR, and RF model design and use to generate one step
ahead prediction.

The obtained results clearly show that the DL/LSTM model outperforms he
SVR and RF ones, in both experiments.

Sample performance table:

Performance of the DL/LSTM model compared to those of SVR and RF models for plant
yield and growth prediction.
 SYSTEM REQUIREMENTS

H/W SYSTEM CONFIGURATION:-

 Processor - Intel dual core

 Speed - 1.1 Ghz
 RAM - 1 GB(min)
 Hard Disk - 20 GB
 Key Board - Standard Windows Keyboard
 Mouse - Two or Three Button Mouse
 Monitor - SVGA

S/W SYSTEM CONFIGURATION

 Operating System : Windows XP/7

 Application Server : Tomcat5.0/6.X/8.X

 Front End : python Tkinter

 Database Connectivity : Mysql.