Research on Intelligent Planting Optimization of Soil Environment
Based on Machine Learning Method
Yu Cao
School of Computer and Information Engineering, Henan University of Economics and Law, Zhengzhou, Henan Province,
450000, China
Keywords: Multiple Logistic Regression, Random Forest Model, Support Vector Machine Model, Machine Learning,
Intelligent Planting
Abstract: Crops not only provide nutrients for human growth but also are the raw materials for many industries. There
are a lot of food supply and food security issues. Crops play an important role in people's lives. This study
trained an intelligent model to judge the high-yield crops in this region based on the content of some elements
in the soil and the characteristics of temperature and humidity. Training multiple logistic regression, random
forest, and support vector classification (SVC) models provide an initial model for this investigation. The
parameters of this initial model were optimized and adjusted to find the best model based on the soil intelligent
planting problem. Overall accuracy, precision, and recall were selected as the evaluation indexes of the three
models. The overfitting of the three models is alleviated by grid search cross-validation. Finally, it is found
that the accuracy of the SVC algorithm reaches 0.99 on both the training set and the test set. This model has
outstanding performance in intelligent planting problems. Training the smart planting model based on the soil
environment can help farmers better select the best crops in different soils to achieve high yields. Dramatic
improvements in land use efficiency and yield could ease today's global food supply problems. The
subsequent use of more accurate climate mitigation data collection equipment is expected to train an
intelligent planting model that combines land mitigation and climate environment.
1 INTRODUCTION
Crops provide people with essential nutrients such as
carbohydrates, proteins, fats, vitamins, and minerals.
The world's population, estimated at 7.7 billion,
consumes 2.5 billion tons of food annually, the
majority of which is derived from crops. Crops are
also important raw materials for various industries.
Crops can provide different kinds of raw materials for
textile, paper, chemical, pharmaceutical, energy, and
other industrial sectors, such as cotton, linen, bamboo,
wood, oil, sugar, starch, cellulose, ethanol, and so on.
Companies can create significant economic value by
further processing crops (Qikai, 2016). However,
global food supply and security are threatened by
multiple factors, such as climate change, geopolitical
conflicts, the COVID-19 pandemic, and volatile food
and fuel prices. The crisis has put hundreds of
millions of people around the world at risk of hunger
and malnutrition, especially in some low-income
countries and regions (World Agric, 2022).
According to the United Nations World Food
Programme, around 345 million people in 82
countries are currently facing severe hunger (World
Agric, 2019). On the other hand, with the
improvement of medical technology, the death rate of
the population is decreasing year by year, and the
population is constantly growing. This factor has also
led to an increasing number of people, and the
phenomenon of food shortage is increasing. With the
advancement of data acquisition technology, people
can collect the desired data through various sensors
(Fengshao, 2023). AI data-based machine learning
algorithms have also been improved. These
algorithms have been widely used in finance, biology,
chemistry, and other fields. In agriculture, many
researchers have made a lot of crop yield prediction
techniques based on climate characteristics,
environmental conditions, and other characteristics.
For example, Lontsi Saadio Cedric et al. collected
climate data, weather data, agricultural yield, and
chemical data in West Africa. They used this data to
build their system based on decision trees, multiple
logistic regression, and K-nearest neighbor models.
Cao, Y.
Research on Intelligent Planting Optimization of Soil Environment Based on Machine Learning Method.
DOI: 10.5220/0012824000004547
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 1st International Conference on Data Science and Engineering (ICDSE 2024), pages 481-487
ISBN: 978-989-758-690-3
Proceedings Copyright © 2024 by SCITEPRESS – Science and Technology Publications, Lda.
481

They then adjusted the hyperparameters to obtain a
model that did not overfit to predict the yield of six
local crops (Saadio et al, 2022). He collected satellite
meteorological data from 98 districts and counties in
Jilin and Shanxi provinces. Based on these data, he
used regression algorithms from random forests and
support vector machines, as well as deep learning
neural network models to predict local corn yields.
Ultimately, each model's prediction effect is assessed
and examined using the mean square error, mean
absolute percentage error, and root mean square error.
The study confirms that the neural network model is
the most effective one, and it also demonstrates the
viability of machine learning techniques in the
agricultural sector (Junxiu, 2021). There is not much
research on machine learning classification
algorithms in agriculture, but these algorithms can use
satellite images or images collected by aircraft to
classify different crops in farmland. This helps to
monitor the health of the field and to manage
irrigation and fertilization promptly. In this study,
based on the classification algorithm of machine
learning, a model is designed to judge which crops
can produce high yields in a given soil environment
based on soil conditions. This model not only
improves the efficiency of land use but also produces
more food. The application of this model can help
alleviate the problem of food shortage.
This study will use three machine learning
methods, which are multiple logistic regression,
support vector machine classification, and random
forest. This study will first train the model of each
algorithm separately based on the data. Secondly, the
optimal model of each algorithm is obtained by cross-
verifying the selection of hyperparameters. Finally,
the overall accuracy, precision, recall, and other
indicators of the best model of each algorithm are
compared to get the best model in this study. This
paper aims to investigate machine learning-based
intelligent planting optimization of soil environment.
2 METHOD
2.1 Data Set Description
Table 1 shows some of the data features and labels in
the data set. This study is based on soil nitrogen
content, phosphorus content, potassium content,
temperature, humidity, soil pH, rainfall data, and
high-yield crop categories. To make the subsequent
data analysis more beneficial, all table data will be
reserved for 3 decimal places. This data comes from
Kaggle. Kaggle is an online platform for data
scientists, machine learning engineers, and other
professionals in data-related fields.
Nitrogen: One of the basic elements of plants, the
main component of chlorophyll. Chlorophyll is the
compound that enables plants to photosynthesize
(Yanxiao et al, 2022).
Phosphorus: Phosphorus is essential for both the
growth of new tissues and cell division. In plants,
phosphorus is also involved in the intricate process of
energy conversion. When phosphorus is added to
soils with low levels of accessible phosphorus, it can
encourage tillering, strengthen roots against cold
stress, and frequently speed up ripening (Miaomiao et
al, 2023).
Potassium: A nutrient obtained by plants from soil
and fertilizers. Potassium can improve disease
resistance, stem strength, drought tolerance, and
winter survival of plants (Hashim & Mohammed,
2023).
Soil temperature: The average soil temperature for
biological activity is between 50 and 75 degrees
Fahrenheit, similar to the temperature of the human
body. These values are conducive to the normal life
functions of the Earth's biota, such as decomposing
organic matter, mineralizing nitrogen, absorbing
soluble substances, and metabolism (Huiyong et al,
2008).
pH: The pH value indicates the acidity or
alkalinity of the soil. The pH range of 5.5 to 6.5 is
ideal for plant growth, as this range guarantees the
availability of nutrients (Chenxiao, 2022).
Rainfall: Rainfall also affects how fast crops grow
from seed, including when they are harvested.
Balanced rainfall and proper irrigation can accelerate
plant growth, which can reduce germination times
and the time between planting and harvesting
(Podzikowski et al, 2023).
Table 1: Partial data on factors affecting crop growth and corresponding high-yielding plants.
Nitrogen Phosphorus Potassium temperature humidity ph rainfall label
0 90 42 43
20.879 82.002 6.502 202.935 rice
1 85 58 41
21.770 80.319 7.038 226.655 rice
2 60 55 44
23.004 82.320 7.840 263.964 rice
3 74 35 40
26.491 80.158 6.980 242.864 rice
4 78 42 42
20.130 81.604 7.628 262.717 rice
ICDSE 2024 - International Conference on Data Science and Engineering
482

Table 2: Statistical overview of factors affecting crop growth.
Nitro
g
en Phos
p
horus Potassium tem
p
erature humidit
y
p
h rainfall
mean 50.551 53.362 48.149 25.616 71.481 6.469 103.463
st
d
36.917 32.985 50.647 5.063 22.263 0.773 54.958
min 0 5 5 8.825 14.258 3.504 20.211
25% 21 28 20 22.769 60.261 5.971 64.551
50% 37 51 32 25.598 80.473 6.425 94.867
75% 84.25 68 49 28.561 89.948 6.923 124.267
max 140 145 205 43.675 99.981 9.935 298.560
2.2 Feature Engineering
Table 2 shows a statistical summary of the factors
affecting crop growth. The three elements Nitrogen,
Phosphorus, and Potassium are necessary for the
growth of crops. The content of each element required
by plants has a large span and significant standard
deviation, which can indicate that the amount
required by each type of plant for these three elements
is significantly different. Most of the temperatures
required for crop productivity are concentrated
around 25, with only a few crops at some extreme
temperatures. The standard deviation of pH in the soil
is the smallest. The pH required for high yield of
crops is concentrated around neutral and slightly
acidic. The difference between humidity and rainfall
is also relatively large, which is also conducive to our
division and practical application of crops. Because
the climate is different around the world, the rainfall
is also different, and different areas are suitable for
different high-yielding crops.
Pearson's formula is used to study the multiple
correlations among the parameters (Layeb, 2023).
Figure 1 shows a hotspot map of correlations between
features. There was a strong positive correlation
between the content of Phosphorus and Potassium,
which made different types of crops produce high
yields. The above situation shows that the content of
Phosphorus and Potassium in an environment of high
yield should increase and decrease consistently. The
Nitrogen element shows a certain negative correlation
with the above two elements. Other features have
relatively small dependencies.
Figure 1. Pearson correlation between features (Photo/Picture credit: Original).
Research on Intelligent Planting Optimization of Soil Environment Based on Machine Learning Method
483

2.3 Proposed a Soil Environment
Intelligent Planting Model Based on
Machine Learning
2.3.1 Multiple Logistic Regression (LR)
A statistical technique for binary classification that
forecasts the likelihood that a result will fall into one
of two categories is called logistic regression.
Multivariate categorization can also be accomplished
with logistic regression. An important technique in
statistics and machine learning, it simulates the link
between the logarithmic probabilities of independent
and dependent variables. (Leijia, 2023).
For each category j, the weight of each feature is
represented by formula (1), The number of features is
represented by the superscript, the number of
categories by the subscript, the total number of
features by n, and the total number of categories by
N.
𝑍

=𝑤


𝑥

+𝑤


𝑥

+𝑤


𝑥

+...+𝑤


𝑥

j = 1,...,N
(1)
The probability that the classification result is a
class for the input point x is given by formula. (2).
P（y=j|Z

） =



∑





(2)
Create training and test sets from the data set.
Thirty percent of the whole data set is the test set,
while the remaining seventy percent is the training
set. Each category of data in the training set will be
calculated by formula (1) and formula (2) to get its
probability in each category, and the maximum
probability will be selected as its final predicted
value. The model's performance is assessed using the
cross-entropy loss function (formula (3)), which
converges to a minimum value. Using gradient
descent algorithm to reduce the loss function. Finally,
the best multiple logistic regression classification
model is obtained.
J(w) = −


∑
y
(

)
log

h

x
(

)


+1−


y
(

)
log

1−h

x
(

)


 (3)
2.3.2 Random Forest Model (RF)
Compared to a single decision tree, this potent tree-
based ensemble learning system outperforms it. By
using retractable sampling to generate multiple
different training subsets, we can train an independent
decision tree for each subset. We can then combine
these decision trees into an integrated model, and for
the prediction task, have the trees vote to get the final
prediction result. To make the decision tree in tree
integration more diverse, we can further introduce
randomness in the selection of partition features for
each node. To be more precise, we can choose a
subset of size K (K less than N) at random from all N
features at each node. From this subset, we can choose
the feature that has the maximum information gain to
be the partitioning feature. (Ruiyan, 2022).
Steps to create a random forest:
Given a training set size N and the number of trees
to be created n, for each tree, N samples are taken
from the total sample to form the training set of the
tree. Repeat n times to form n numbers to form the
forest
For M features of each sample, m features are
selected as the training features (m < M).
Use these m features to make the tree extend as far
as possible
Enter the new data you want to predict into n trees
and vote for the result.
An indicator of a data set's purity is entropy.
Higher entropy indicates greater impurity, whereas
lower entropy indicates higher purity. When building
a decision tree, we can use entropy to decide which
feature should be divided at the node. In general, we
choose the feature that will reduce entropy the most.
This means that you want to make the subset as pure
as possible by splitting it. In classification problems,
there are two formulas for calculating entropy:
information entropy (formula (4)), and Gini impurity
(formula (5)).
H(D) =
∑
p

log

p

(4)
H(D) =
∑
p

(1 − P

)

(5)
2.3.3 Support Vector Machine Model (SVM)
Finding a hyperplane to divide the sample into
segments is the goal of the binary classification
algorithm known as support vector machine, where
the idea behind segmentation is to maximize the
interval. Let the hyperplane be 𝜔

𝑥+𝑏=0, given a
sample x, 𝜔

𝑥+𝑏 Represents the distance to the
hyperplane. Suppose that after extending the space
point to N-dimensional space, the distance from the
point to the hyperplane is formula (6) (Saadio et al,
2022):
d=
∗|

|
||||
(6)
According to the Lagrange multiplier method,
KKT condition, and duality problem, we can calculate
a set of w values of the optimal SVM objective
function according to the following steps.
ICDSE 2024 - International Conference on Data Science and Engineering
484

min


||ω||

s.t 1 − y ∗(𝜔

𝑥+𝑏)≤0 (7)
Since the original objective function is a
downward convex function, according to the duality
problem, we can see that a) must be a strong duality
problem, which can be converted to
L(w,b,λ) =


||ω||

+
∑
λ

[1 − y

(


𝜔

𝑥

+𝑏)]
s.t λ

≥0 (8)
The process of building an SVM model consists
of three stages: feature selection, feature
transformation, and hyperplane creation
3 RESULTS AND DISCUSSION
3.1 Model Evaluation
In this investigation, the model's quality was assessed
using three indicators: recall, accuracy, and precision.
The training of the model on the training set and the
verification on the test set can well judge the accuracy
of the training model and whether there is overfitting
and generalization ability through these three indexes
Table 3: Three model evaluation indicators using default parameters.
Accuracy Precision Recall
Model Train Test Train Test Train Test
LR 0.975 0.948 0.976 0.951 0.976 0.950
RF 1.000 0.995 1.000 0.995 1.000 0.995
SVC 0.975 0.978 0.978 0.982 0.978 0.981
In the case of the logistic regression model, there
is an overfitting issue and an excessive variance
between the accuracy on the training and test
sets.When compared to the other two models, its
prediction accuracy is not as good, and its precision
and recall rate are also evident. For the random forest
model, its score on the training set has reached an
astonishing 1, and there is a high possibility of
overfitting. However, the results on the test set are
very good, indicating that after some adjustments and
optimization, the model has a great accuracy to solve
this problem. The SVC algorithm's training set score
and test set score are fairly similar, and the features
acquired in the training set are practically applicable.
Hyperparameter selection and cross-validation are
employed to address the overfitting issue and improve
the model training's predictive accuracy in order to
address the aforementioned issues.
3.2 Training Optimization
To ensure that the model does not learn too much
from the training set. The author uses grid search
cross-validation to identify the model's ideal
hyperparameters. The grid search is a parameter
tuning technique where the parameters are adjusted
within a specified range with a defined step size.
These modified parameters are used to train the
learner, and the parameter that produces the best
accuracy on the validation set is chosen. This process
involves training and comparing multiple models.
Using the GridsearchCV library, takes the machine
learning model, the hyperparameter selection grid,
and the number of cross-validations as inputs. The
best estimator (machine learning model) of the best
parameters is given.
The maximum number of method convergence
iterations and the reciprocal of the regularization
coefficient are the linked hyperparameters for the
logistic regression algorithm. The regularization
coefficient can assist the model in overcoming
overfitting, significantly increase the model's
accuracy when applied to unknown data, and
streamline the intricate model by getting rid of
mistakes and noise. If the number of algorithm
iterations falls within a suitable range, the model's
efficiency can be raised. After using grid search and
cross-validation, the best model parameters of the
multiple logistic regression algorithm in solving this
problem are {C = 0.01, max_iter = 700, cv = 5}. The
details of the model are shown in Table (4).
For the random forest algorithm, the associated
hyperparameter is selected as the number of the
number in the created forest.The quantity of
appropriate trees in the forest makes sure the model
picks up the features of the data set without becoming
overfit. After using grid search and cross-validation,
the best model parameters of the multiple logistic
regression algorithm in solving this problem are
{'max_depth'= 6, 'n_estimators'= 100, cv = 5}. The
details of the model are shown in Table (5).
Research on Intelligent Planting Optimization of Soil Environment Based on Machine Learning Method
485

Table 4: Multiple logistic regression to optimize details.
mean
_
fit
_
time
p
aram
_
C
p
aram
_
max
_
iter
p
arams mean
_
test
_
score rank
_
test
_
score
0.3864604 0.01 700
{'C': 0.01, 'max_iter':
700}
0.977 1
Table 5: The random forest algorithm optimizes details
mean
_
fit
_
time
p
aram
_
max
_
de
p
th
p
aram
_
n
_
estimators
p
arams mean
_
test
_
score rank
_
test
_
score
0.190494919 6 100
{'max_depth': 6,
'n_estimators':
100}
0.990 1
For the SVM classification algorithm, the related
hyperparameters are the penalty coefficient C of the
error term and the type of kernel function. The size of
C has a strong correlation with the accuracy and
generalization ability of training results Choosing the
appropriate penalty coefficient has a great influence
on the fitting ability and generalization ability of the
model. After using grid search and cross-validation,
the best model parameters of multiple logistic
regression algorithms in solving this problem are
{'C'= 10, 'kernel'= 'rbf', gamma='auto',
probability=True}. The details of the model are
shown in Table (6).
Table 6: Support vector machine classification optimization details.
mean
_
fit
_
time
p
aram
_
C
p
aram
_
kernel
p
arams mean
_
test
_
score rank
_
test
_
score
0.12366333 10 rbf {'C': 10, 'kernel': 'rbf'} 0.985 1
Table (7) shows the accuracy rate, precision rate,
and recall rate of the three optimized models after
training. Compared with the default parameters
before optimization, the data detection on the test set
is more accurate. Furthermore, there is very little
score difference between the training and test sets,
and the overfitting issue has also been resolved. The
LR model has an overall accuracy of approximately
0.98, the RF model has an overall accuracy of
approximately 0.99, and the SVC model has an
overall accuracy of approximately 0.99. The model
construction of the three algorithms for this problem
is relatively good, but by comparing the time of model
fitting in Tables 4, 5 and 6, Out of the three
approaches, the support vector machine's
classification model is the most effective. In general,
all three models show their ability to solve this
problem after being optimized.
Table 7: Three model evaluation indexes using optimization parameters
Accurac
y
Precision Recall
Model Train Test Train Test Train Test
LR 0.989 0.983 0.989 0.983 0.989 0.983
RF 0.989 0.990 0.989 0.991 0.989 0.991
SVC 0.989 0.992 0.989 0.992 0.989 0.992
4 CONCLUSION
This study is based on soil nitrogen content,
phosphorus content, potassium content, temperature,
humidity, soil pH, and rainfall data to intelligently
select the most suitable for high-yield crops in a
specific environment. The data in the data set is
preprocessed first to make the formation of the model
faster. Second, to find the association between the
features and hidden information inside the data,
feature engineering is used to the data set. Three
models were employed in this study to train the data
in an initial stage. It is evident from the experimental
data that the three models are not very convincing in
the training and test sets and have certain flaws.
Finally, a more comprehensive model for this
problem is generated by applying the grid search
cross-validation method to optimize the above model.
The SVC model has a considerable advantage in
fitting time and not only has an accuracy rate of 0.99
in the validation of the test set, but it also has a
precision and recall rate that are close to 0.99. Due to
the large amount of data and the significant
ICDSE 2024 - International Conference on Data Science and Engineering
486

classification of each crop feature, the three models
have significant performance in the three indicators
after the final optimization. The difference in overall
accuracy lies in the difference in the division of
individual crops. The SVC model has a more
significant effect than the above two models on the
environmental high yield of aconitum bean.
Machine learning is more and more widely used
in today's life, and many researchers in agriculture
have conducted in-depth research on the yield
prediction of a single or several crops. This research
can help people to know what kind of soil in their
country or region is suitable for growing crops. This
practice can prevent the loss of yield and income
caused by random planting, and help people to
maximize their profits in farming. If some areas are
not suitable for planting some crops, the soil
environment can be adjusted according to the range
values of various characteristics in the above study to
achieve the purpose. This study can also be combined
with the other yield models described above to not
only identify the best crops to grow in the area, but
also to predict the yield. In future development, the
model can be further improved to add more
characteristics such as climate, the region where the
soil is located, the amount of sunshine and other
atmospheric environments. Combining the soil
environment with the atmosphere further improves
the smart planting model, allowing more accurate
classification of which crops are more productive in a
given region.
REFERENCES
Y. Qikai, the Ningxia Hui Autonomous Region, Ningxia
Southern Chemical Technol. Co. (2016).
International food and agriculture trends, World Agric.
(05):119-121 (2022).
Global Food Crisis Report, World Agric. (05): 96 (2019).
Z. Fengshao, Wirel. Interconn. Technol. 19(02): 109-111
(2023).
L. C. Saadio, Y. W. A. Hamilton, A. Rubby, et al, Smart
Agric. Technol. 2 (2022).
H. Junxiu, Jilin Agric. Univ. (2021).
F. Yanxiao, J. Yongda, L. Weiwei, For. Sci. Technol. Inf.
54(02): 109-111 (2022).
C. Miaomiao, Y. Bin, R Guangqian, et al, Flora. 309 (2023).
B. A. Hashim, H. A. Mohammed, IOP Conf. Ser. Earth
Environ. Sci. 1262(8) (2023).
Y. Huiyong, L. Meixue, S. Jieya, et al, Priv. Sci. Technol.
(03):102 (2008).
G. Chenxiao. Shanxi Agric. Univ. Agricultural University
(2022).
L. Y. Podzikowski, M. M. Heffernan, J. D. Bever, Front.
Ecol. Evol. 11(2023)
A. Layeb, Int. J. Intell. Syst. Appl. 15(2):37-42 (2023).
L. Leijia, Southwest Univ. (2023).
L. Ruiyan, Nanfang Agric. Machinery, 53(22):63-65+87
(2022).
Research on Intelligent Planting Optimization of Soil Environment Based on Machine Learning Method
487