Journal of Applied Sciences, Information and Computing (JASIC) Pages (25-33), 2023

Journal of Applied Sciences, Information and Computing

Volume 4, Issue 1, July 2023
School of Mathematics and Computing, Kampala International University

ISSN: 1813-3509 https://doi.org/10.59568/JASIC-2023-4-1-03

DEVELOPMENT OF A MACHINE LEARNING REGRESSION MODEL FOR

ACCURATE SUGARCANE CROP YIELD PREDICTION, JINJA – UGANDA

Yuma Erick1, Chinecheremu Umezuruike2, Nasasira Jossy3, Balyejusa Gusite4

1
Directorate of Research, Innovation, Constance and Extensions, Kampala International University,
Uganda. [email protected]
2
Assistant Professor Software Engineering, Bowen University, [email protected]
3
Department Of Information Technology, Kampala International University, Uganda.
[email protected]
4
Department Of Information Systems, Kampala International University, Uganda.
[email protected]
Abstract
Sugarcane is one of the key crops grown worldwide and used for sugar processing, food, alcohol,
biogas, fertilizer, and other products. There is a problem with Sugarcane yield prediction, yields
aren’t accurately predicted, and this creates an impact on yields. This research looks at identifying
methods used for the prediction, design, development, and evaluation of the three machine-
learning regression models used for predicting sugarcane yields in Uganda. This research was
implemented using Data Science methodology, several machine learning algorithms for prediction
of yields on dataset have been analyzed. The collected and analyzed dataset in this research had
one output/ dependent variable and eight independent variables. The algorithms used to develop
the prediction models are the Multiple Linear Regression algorithm, Decision Tree Regression
algorithm, and Random Forest Regression algorithm to predict the output. The dataset of 3 years,
2019, 2020, and 2022 was considered and merged to train and test the model at a ratio of 80% to
20%. The accuracies of the individual models were compared after training, testing the dataset,
and evaluation. The multiple Linear regression model results indicate that out of 100%, the model
accuracy was 76.5%, the Decision Tree Regression Model scored 89.2%, Random Forest
Regression Model was 94.6%. The random forest model came out as the best model. The Random
Forest model has a percentage improvement of 60.4%. In future research, researchers can work

25
https://doi.org/10.59568/JASIC-2023-4-1-03
 Journal of Applied Sciences, Information and Computing (JASIC) Pages (25-33), 2023

on, A web-based machine learning model, Deep learning methods used to improve the model and
more data can be used to improve the accuracy.
Keywords: Machine Learning, Kakira Sugar Limited, Random Forest Regression, Decision
Tree Regression, Machine Learning Algorithm, Multiple Linear Regression

1. INTRODUCTION pesticides and fertilizers to ensure healthy and high-

yield sugarcane.
Agricultural Organization in UN stated that 90 and
The problem faced by most farmers growing
above countries grow sugarcanes in an area of 26
sugarcane is that they don’t follow the weather
million hectares and 1.8 billion tons of global
patterns and soil structure before choosing where to
harvests of sugarcane (Taherei Ghazvinei et al.,
plant the crop. They don’t invest time and money to
2018). Brazil is the leading sugar producer in the
research the conditions favorable to bring out good
world with 42 million metric tons in 2020/2021
yields in the area they want to produce the crop.
mounting to 179.66 million metric tons worldwide,
followed by India, China, Cuba, Pakistan, Mexico,
Iran, and Thailand (Vijayakumar & Bozward, 2021,
Shahbandeh, 2021). 1.2 Sugarcane harvest
Uganda is the leading sugarcane grower in East Sugarcane at times doesn’t need replanting because
Africa, followed by Kenya and Tanzania. The it’s the top of the plant which is removed while
agricultural sector is one of the vital parts of the harvesting. It is harvested either manually or
economy in Uganda because millions of people are mechanized. Manual harvest methods can easily
engaged in it. Sugarcane is one of the main crops make the sugarcane gardens to be burnt. All the
grown in Uganda and is a significant contributor to leaves are removed so that sugarcane is then
the country’s economy. In 2020, sugarcane manually cut to the ground then removed and taken
production for Uganda was 5.78 million tonnes to the mill by tractors or other means of transport.
growing at an average annual rate of 3.76%. The mechanized harvest uses machines to extract the
sugarcane as it moves in the garden, while it is loaded
1.1 Sugarcane Production into a truck. The mechanized harvesting method is
Sugarcane farmers have multiple fields with canes at seen as the future of cane harvesting. (Spencer,
different stages of growth to make sure there is February 21, 2020).
always availability of sugarcane to supply the mill.
The cane is planted and all through to harvest, the Sugarcane also called Saccharum officinarum is a
process can take from 8 months to 18 months in the perennial species of tall grass of Asian native whose
fields. This makes sugarcane a slow perennial crop growth has lasted for more than 4000 years planted
to produce. Preparation of sugarcanes before for the juice where sugar processing is done and
planting is very vital. The cane leaves are removed doesn’t need to be re-planted every year. When
and therefore cut into 20cm pieces. The field is harvesting, the stalk is cut above the root level so that
prepared with incisions made at approximately 15cm it can regrow. In estimation, the plant crop is
to 20cm in width. The pieces of sugarcane are placed expected to yield 100%, Ratoon 1 – 70%, Ratoon 2 –
in the horizontal holes, then covered with soil and left 60%, and Ratoon 3 – 50% tons for Kakira Sugar
to grow. The production of cane starts with healthy Limited. The plant height reaches 3–6 meters, it is
soil. This is done by adding nutrients and water. It also used for direct consumption as food. The world’s
can take 12 to 18 months approximately to mature. most sugar cane is cultivated in subtropical and
During this time, the cane must be treated with tropical places. Biofuel production, alcohol,

26
https://doi.org/10.59568/JASIC-2023-4-1-03
 Journal of Applied Sciences, Information and Computing (JASIC) Pages (25-33), 2023
fertilizer, and electricity generation are also obtained Rainfall, temperature, humidity, floods, and others.
from sugarcane. (ii) soil structure for example; loam soil and sandy
Sugarcane manufacturing methods have been soil. (iii) sugarcane variety. (iv) Distance of
designed in India by 400 BC. By the 11th century plantations. (v) planting area. (vi) Age of sugarcane.
AD, sugar was introduced to Europeans during the (vii) irrigation and drainage system. (viii) Pest and
Crusades and then taken all over the European diseases, (ix) Fertilizers and fires. (x) method of
countries, the gentleman called Christopher harvesting and others. (Herbet, 2020)
Columbus is said to have imported sugarcane to the
West Indies and more than 75% of world’s sugar is These cause a problem of fluctuation and inaccurate
got from sugarcane today (Sugarcane Profile, 2021). yields of sugarcanes which brings about low sugar
Sugarcane production in Uganda started in the 1920s prices, changing production trends of sugar, high cost
in the Busoga region introduced by Muljibhai of production, high prices payable to farmers, high
Madhivan with 320 hac in Kakira sub-county in cost of Production, use of unreliable and old
Jinja. sugarcane growth regions in Busoga (KSL) machinery, mill sizes, fluctuating prices of canes,
with 10,000 hectares of nucleus estate 856 hectares inaccurate weeding and harvesting timing, financial
of satellite estates and over 4,400 hectares at crisis. The problem brought by the inaccurate
Kayunga estate. prediction methods used cause difficulty in getting
the accurate results of sugarcane yield. The failure to
get an accurate prediction of sugarcane yield

The main manufacturers of sugar in Uganda are leads to challenges in planning, decision-making,
Kakira Sugar Works, Kinyara Sugar Works Limited, and budgeting for resources among the sugarcane
GM Sugar Uganda Limited, Sango Bay Estates farmers, factories, and organizations.
Limited, Sugar Corporation of Uganda Limited and
others include the following among others: Amuru 1.4 Justifications of the Study
Sugar Works Limited, Atiak Sugar Factory, Bugiri ML models have successfully been implemented in
Sugar Factory, Busia Sugar Limited, Hoima Sugar the prediction of crop yields around the world, the
Limited, Kamuli Sugar Limited, Kenlon Industries algorithms modeled include; MLR, RF, Decision
Uganda Limited, Mayuge Sugar Industries Limited, Tree, K Nearest Neighbor, neural networks, Logistic
Mukwano Sugar Factory, Seven Star Sugar Limited, regression, support Vector Machine and K-means,
Sezibwa Sugar Limited, Buikwe Sugar Works and Convolutional Neural Networks, Recurrent
Limited, Sugar & Allied Industries Limited, Uganda Neural Networks and interaction-based models
Farmers Crop Industries Limited. among others.
1.3 Problem Statement Sugarcane crop yield prediction is highly significant
In Uganda, there is a problem with Sugarcane crop since it provides insights that guide companies,
yield prediction. The yields are not accurately industries, and organizations in improving
predicted and this creates an impact on sugarcane informative knowledge to the stakeholders for mill
yield management. management, sugarcane cultivation monitoring,
sugar production monitoring, informed economic
At Kakira Sugar Limited Agronomy Section, there is and management decisions, planting and crushing
a problem with sugarcane crop yield prediction. They season, corruption, availability and balance of
have challenges related to sugarcane growing and fertilizer.
sugar production due to that. There are reasons for Research shows the relevance of machine learning
the problems of yield fluctuation in the last decade and indicates that it provides predictions that are
and there are factors responsible. Numerous factors reasonable with higher flexibility and faster results.
are in play and affect sugarcane crop yields in several aspects are to be investigated concerning
Uganda and this has an impact on the final yields on accuracy prediction, time taken to predict, supply
maturity. They include; (i) weather for example;
27
https://doi.org/10.59568/JASIC-2023-4-1-03
 Journal of Applied Sciences, Information and Computing (JASIC) Pages (25-33), 2023
management, human capital, and machinery using partial swarm and other techniques to improve
management among others. These organizations its accuracy (Sunil Kumar, 2019).
and companies have engaged a manual approach to
yield prediction, which has not proven to be Forecasting corn yield with ML ensembles was
effective and accurate. Hence, there arises a need for done in three US corn Belt states Indiana, Illinois,
accurate sugarcane yield prediction using Machine and Iowa with datasets that comprise environmental
learning. data like weather, soil, and features of management
of the 2 distinct scenarios; season partial
1.5 Related Literatures. knowledge, in-season weather complete knowledge
Dimo Dimova looked at sugarcane yield estimation and 3 scales; agricultural district, county and state
through remote sensing time series and phenology level. Historical yields of corn at the county level
metrics. The existing regression-based crop yield from USDA National Agricultural Statistics were
estimation approaches used before relying on sets of obtained from 2000 to 2018. Methods used include;
predictor variables that are specific but have not been LR, Bayesian search, LASSO Regression, XGBoost
compared well. The research compares and and LightGBM, Random Forest, and stacked
demonstrates the use of three sets of object-based Generalization. After prediction was made for the
predictors for sugarcane yields in the platform of ensembles, RL and LASSO regression are the
agricultural monitoring knowledge. This uses earth models whose prediction overestimated true values,
observation data of sentinel-2 satellites captured and others underestimate corn yields. (Shahhosseini
between 2018 and 2019 for an area of 10,000 et al., 2020).
hectares in Ethiopia. Many regression models were The study was done by Mupangwa et al., 2020 to
used. (Dimo Dimova, 2022). answer the following questions. (i) Can ML
Accurate crop prediction was done using the K- techniques predict maize grain yields under
Nearest Neighbor algorithm in Mangalore, Kodagu, conservation agriculture? (ii) How close can ML
Kasarago, and other districts of Karnataka state. The algorithms predict maize grain yields under
real-time environmental parameters like soil type, conservation agriculture-based cropping systems in
Rainfall, humidity, irrigation type, previous yields, highlands and lowlands of Eastern Africa and
location, price, year, type of crop, crop diseases, and Southern Africa. The maize yield dataset for seven or
its symptoms datasets in India were collected. more years was collected from many websites and
Related Data on commonly grown crops in the 80% was used in training and 20% in testing the
region like coconut, Cardamom, coffee, Areca nut, algorithm (Mupangwa et al., 2020).
Ginger, Tea, Paddy, Ground nut, Black gram, Random Forest Algorithm (RF) has been used to
Cashew, and Pepper was collected. (Karthikeya et predict accurate sugarcane yield at Tully, in
al., 2020). Northeastern Australia. Annual variation in
In soft computing, support vector machines have productivity of sugarcane and suitability of predictor
acquired reasonable significance in making variables generated from crop models coupled with
predictions. In this paper, SVM models based on observed climate and climate prediction indices
classification were developed to predict rice yields (Everingham et al., 2016).
in India. The historical Data on rice production for
the years 1950 to 2014 was obtained from the (Wickramasinghe et al., 2021). Looks at modeling the
Directorate of Economics and Statistics, Ministry of relationship between rice yield and climate variables
Agriculture, Indian government. The prediction using statistical and ML techniques. Several climate
which was accurate for the four-year relative variables were put under consideration for the
average increase was achieved as 75.06% using a application of both methods. Regression techniques,
4fold cross-validation Method. The experimental Support Vector Machines, and Artificial Neural
software used for this work was MATLAB. This Networks are the techniques applied in this
work can be improved by redefining training relationship modeling. Data collected included rice
patterns and learning parameters can be optimized
28
https://doi.org/10.59568/JASIC-2023-4-1-03
 Journal of Applied Sciences, Information and Computing (JASIC) Pages (25-33), 2023
harvest, yield, and climate data for 2 districts in Sri Statistical Calculation Using Excel.
Lanka for more than 3 decades. In the statistical method, the area and cycle of canes
are the major factors of yields that are considered.
2. METHODOLOGY The field area is the space the sugarcane plant covers
2.1 Data Science and the cycle of the canes is the number of times the
plant is left to germinate from the stacks after
Data science is a field of study that uses scientific germinating.
methods, processes, systems, and algorithms to
extract insights and knowledge from structured and The area is in hectors and the cycles are called
unstructured data. It looks at varieties of data by ratoons, the first ratoon(R1) gives 100%, R2 gives
using modern tools and techniques to find patterns 70%, and R3 gives 60%.
and derive knowledge and business decisions. There The prediction is made by multiplying the area of the
are more fields under data science, these include; canes and the ratoon of the cycle. The accuracy gives
Artificial Intelligence, Deep Learning, and Machine got is 59% of the sugarcane yields.
Learning. (Logallo, 2019)
That is:
Data science methodology has a routine of finding
solutions to specific problems using a cyclic Prediction = Area * Ratoon of The Crop.

process to guide data scientists which has the 2.3 Design a Predictive Model for Sugarcane
following Steps: Business Understanding, Analytic Yield
Understanding, Data Requirements, Data The design of the predictive model consists of the
Collection, Data Understanding, Data Preparation, steps undertaken to make the model meet the main
Modelling, Evaluation, Deployment, Feedback objective of predicting the sugarcane crop yields as
shown in the conceptual diagram (Framework)
Analysis was performed to eliminate noise and
below.
outliers from the data. The necessary variables to
enable prediction were selected by use of feature
selection methods which included the filter method
which drops features according to the correlation Machine Learning Prediction Model Design
with the output variable and the wrapper method
which split the dataset and trained the model by
adding and subtracting the features.
2.2 Methods Used for Sugarcane Yield Prediction
in Uganda
Observation Method.
In Kakira Sugar Limited, the fields have attached
supervisors who monitor and take care of the
sugarcanes. Since they are in touch with the farmers
and can observe how the sugarcanes grow for
specific gardens. With their observation by looking
at the entire field of the sugarcane, they can tell how
many trucks of sugarcane can be got. They consider
the size of the stems of the sugarcanes and the area Figure 1 Machine Learning Prediction Model
covered. The observation method gives an accuracy Design
of 28% of the yields.

29
https://doi.org/10.59568/JASIC-2023-4-1-03
 Journal of Applied Sciences, Information and Computing (JASIC) Pages (25-33), 2023
Data Collection Model Training
The main variables that affect sugarcane crop yields After preprocessing the collected data, it is split
at Kakira Sugar Limited in Jinja district are into a train set and a test set. The training set is
environment, weather, and management which fed into the developed algorithm. The data was
comprise different components of the mentioned processed by the Algorithm and the model is
variables like soil structure, rainfall, temperature, output which can get the target value in the data
area, and others. Since some of the gardens for provided. The model was trained to develop a
sugarcane in some locations are not irrigated, learned model. The prediction is made by the
irrigation was not taken as a feature to keep the model by providing the other attributes of the
consistency of the dataset. The sugarcane crop yield dataset.
data for three years were obtained from KSL
databases and weather data from the weather station Model Evaluation
at the industry in Kakira sub-county located in Jinja. The model developed is tested to find out if it
The variables are explained below. can predict the target value quickly and
accurately. This can be achieved through model
1. Actual Sugarcane yields for the previous 3 tuning done by optimization of the model
years after harvest. parameters to get the Algorithm’s best
2. Area of plant cane per plot. performance. R2 Score was used to get the
3. Age of the sugarcane. accuracy.
4.Sugarcane crop cycle (PC = 100, R1=80,
R2=60, R3=40) Flow Chart for Sugarcane Prediction Model
5.Temperature in degrees Celsius.
6.Rainfall in millimeters.
7.Soil Potassium.
8.Soil PH.
9.Soil silt.
Data Preprocessing
The following pre-processing operations were
considered on the created dataset to make it ready to
train the proposed ML model.
• Removal of weather features which are
before planting and after harvesting (out of
season).
• Aggregation of the weather dataset.
• Observations with lower yields were Figure 2 Flow Chart for Sugarcane Prediction
considered outliers and were left out of the Model
dataset. The flow chart above shows how the model
• The independent variables were scaled to fit operations are undertaken when predicting the values
in the set range. of the sugarcane yields from the start to the end.
• Feature selection from the dataset. The model starts by importing the necessary libraries
and loading the dataset, preprocessing the data by
Dataset Splitting choosing the most important variables that predict
The dataset used in the ML model was split into the yields of sugarcane, training the model using
2 portions, train set and test set. This was done 80% of the dataset, training the model using 20% of
in the ratio of 80% Training and 20% testing the dataset, and then predicting the model.
sets.

30
https://doi.org/10.59568/JASIC-2023-4-1-03
 Journal of Applied Sciences, Information and Computing (JASIC) Pages (25-33), 2023
2.4 Exploratory Data Analysis • Dataset preview
EDA is an approach used to analyze the datasets to
summarize the main characteristics and patterns with
visual and non-visual methods. It was applied to
know and understand the data, it helped the analyst
to spend less time on coding and focus more on the
data. After data collection, it underwent processing
before EDA was done. The cleaned dataset and
knowledge from EDA were used to perform
modeling and reporting.
Exploratory Data Analysis is as follows;
• Import the necessary required tools and libraries
in data analysis.
Pandas is a Python library used to examine,
manipulate, and clean the tabular data. Figure 3 Dataset preview

• Check the number of entries, data types, and

column types using built-in functions.
Numpy library which offered special arrays which
are different in storage and handling. It requires less The dataset has 844 records which were selected
space and its first in terms of speed. after removing outliers out of 1400 records of data.

Matplotlib library which was used to visualize our • Get the mean, standard deviation, minimum
dataset to know the patterns. values, and maximum values.
• Get insights of the value numbers in each
The script-learn library is a collection of advanced column which give information about null
machine-learning algorithms used by Python. and duplicate data.
Seaborn was used to visualize data based on • Plot graphs to get information about the
matplotlib and also provides a high-level interface variables.
for drawing informative statistical and attractive
data = pd.read_csv (“tryy.csv”) was used to assign
the dataset to the variable data which was used to
train and test the model. 2. Results
Accuracy of the Machine Learning Models Used

Figure 4 Accuracy of the Machine Learning

Models Used

31
https://doi.org/10.59568/JASIC-2023-4-1-03
 Journal of Applied Sciences, Information and Computing (JASIC) Pages (25-33), 2023
Random Forest Machine Learning Model The results showcase the utilities of Random Forest
Predicted Results Regression for sugarcane crop yield prediction.
There are many advantages over traditional ML
Regression and Decision Tree Regression in
prediction. Random Forest regression models have
been shown to perform better than MLR and DTR in
the explanation of variability in the dataset and the
results strongly indicate the case in sugarcane crop
yield prediction. It also has an upper hand when the
predictors of sugarcane yield highly correlate with
each other. RF regression chooses the best variable
after splitting the responses at each node of the trees
and averages the trees of the forest. RF can also use
many types of predictor variables more easily than
Figure 5 Random Forest Machine Learning traditional linear and non-linear regressions. For
Model Predicted Results example: Area (Ha), Age, Rainfall, Temperature,
3. DISCUSSIONS Soil PH, Soil Silt, soil k, and Yield are continuous
while Crop Cycle is categorical.
The results show that RF regression is a highly
effective algorithm for sugarcane crop yield The dataset was not easy to prepare because of many
prediction. The RF model outperformed the ML factors which include, mistrust from the sugarcane
Regression model and Decision Tree Regression company, improper storage of data, noise in the data,
with a higher prediction accuracy. Random forest has many unnecessary features of the data, and many
a value of 94.6% which indicates that the predicted others. This made the research so hard.
and the actual values of the yields are very close to The percentage improvement of the new model is;
each other, Multiple Linear Regression has a value of
76.5% which indicates that the predicted and actual Old model percentage = 59%
values are fairly close to each other and Decision
New model accuracy = 94.6%
Tree Regression model has a value of 89.2% which
indicates that the actual and the predicted values are
close to each other. The models perform better
compared to the observation = (Difference of old and new model accuracy /
old model accuracy) * 100%
= (35.6/59) *100%
method which gives an accuracy of 28% and the
Statistical method which gives 59% accuracy. = 60.3%

Even though RF has widely been used for 4. CONCLUSION

classification in recent studies, to date few studies The project presents a machine learning prediction
have used its regression capability for yield and model for sugarcane crop yield prediction. It allows
productivity studies in ecology and agriculture. The real-time prediction throughout the year, also
results in this research indicate that Random Forest currently applicable in Uganda at Kakira Sugar
Regression is desirable for prediction in the Limited in Jinja.
agricultural field.
The Machine Learning model was developed to
RF algorithm intrinsically separates random subsets predict sugarcane crop yields prediction and
of the dataset to perform testing using the remaining examined several factors that affect model
dataset for training the model. It also provides performance. It provides us with state-of-the-art
important information about the variables and the accuracy in prediction and shall have a great role in
dependence. sugarcane production. Data science methodology
32
https://doi.org/10.59568/JASIC-2023-4-1-03
 Journal of Applied Sciences, Information and Computing (JASIC) Pages (25-33), 2023
was used and the design of the models was done. [3] Marsland, S. (2009). machine learning An
Three Machine learning algorithms were used to Algorithm Perspective. Massey University.
implement the model Design with the sugarcane
dataset. The three regression models have a good [4] S R Krishan Priya a, R. K. (2022). Sugarcane
performance on the dataset provided to them. Out of yield forecast using weather based discriminant
them, Random Forest has the best performance with analysis. ELSEVIER B. V, 4.
an accuracy of 94.6% which can make an accurate
prediction of the sugarcane yield. The prediction
model can be used in sugarcane industries, [5] Spencer, R. (February 21, 2020). The Sugar
organizations, and the farmers of sugarcane. The Series : Sugar Cane Production – Growing,
percentage improvement of the new model is 60.3% Harvesting, Processing and Refinement.
compared to the statistical method. CZARNIKOW.

5. REFERENCES
[1] Dimo Dimova, J. H. (2022). Sugarcane yield [6] Terence Epule Epule1, J. D. (2018). The
estimation through remote sensing time series and. determinants of crop yields in Uganda: what is the
ELSEVIER B. V, 13. role of climatic and non-climatic factors? Agriculture
& Food Security.
[2] Herbet, I. (2020). Prediction of sugarcane at
Kakira Sugar Limited. Logallo, N. (2019). Data
Science Methodology 101. Towards Data Science.

33
https://doi.org/10.59568/JASIC-2023-4-1-03