A Machine Learning Approach to Rainfall

Prediction in Bangladesh
Md. Fazle Rabby
Department of CSE,
Daffodil International University
Savar, Dhaka, Bangladesh.
rabby15-5252@[Link] and food security. They nurture crops and
replenish water supplies. However, the same
rains that keep Bangladesh afloat can also cause
Abstract:- Bangladesh's socioeconomic terrible floods that uproot populations and
landscape is significantly impacted by endanger livelihoods. Rainfall is one of the most
rainfall, which has an effect on livelihoods, common natural disasters in Bangladesh which
agriculture, and disaster management. Using rigorously affect agro-based economy and
machine learning (ML) to create a rain people’s livelihood almost every year [1]. The
prediction model, this study recognizes the monsoon rainfall has very important affect on the
potential of artificial intelligence (AI) in agricultural production, livestock as well as
tackling these issues. Six machine learning human ecology [2]. Being able to predict rainfall
(ML) models—Logistic regression, K- with accuracy is not just a scientific endeavor but
also a vital instrument for safeguarding the
Nearest Neighbors (KNN), Decision trees,
welfare of millions of people.
Support Vector Machines (SVM), Random
Predicting the onset, intensity, and duration of
Forest, and Ensemble classifiers—were rainfall depends on an understanding of the
trained and assessed using a dataset of 3272 complex dance of monsoon systems, the impact
records gathered from the Bangladesh of the beautiful Himalayas rising to the north, and
Meteorological Department (BMD) website. the constant threat of climate change on weather
With an amazing accuracy of 83.66% and patterns. Traditional forecasting techniques are
better performance in terms of accuracy and useful, but they frequently fall short in capturing
F1 Score, ensemble classifiers stood out as the subtleties of these dynamic elements, which
the most accurate predictor among them. This limits accuracy, especially for short-term or
study highlights how crucial it is to use AI localized projections.
With the introduction of machine learning (ML),
technology to improve Bangladesh's rainfall
a new era begins. This potent technology presents
forecasting, with important ramifications for
a ground-breaking method for predicting
both agricultural sustainability and disaster Bangladesh's rains. Envision extensive historical
preparedness. meteorological
data, including pressure, temperature, humidity,
Keywords: Rainfall, Machine learning, wind patterns, and even satellite imagery, being
Ensemble classifiers, climate, Agricultural painstakingly examined by advanced algorithms.
sustainability Compared to traditional methods, these
algorithms have the potential to unleash a new
I. INTRODUCTION level of precision and localization in rainfall
Bangladesh, a country with lush plains and forecasts because of their capacity to recognize
powerful rivers slicing through it, depends on the intricate relationships within this data.
monsoon rains to survive. Predictable patterns of The use of machine learning techniques for
precipitation are essential to the nation's rainfall prediction in Bangladesh is examined in
agriculture, which is the backbone of its economy this research article. We set out to investigate a
range of machine learning models, carefully
evaluate each one's efficacy, and shed light on the
advantages and disadvantages of applying this patterns on agriculture are significant in
technology to rainfall forecasting in this area. The Bangladesh.
results of this study have the potential to greatly
Rahee et al. [6] contributed to understanding of
enhance tactics for disaster preparedness,
rainfall dynamics in Bangladesh and provides
agricultural planning, and the management of
valuable tools for forecasting and managing
water resources. It is impossible to overestimate
agricultural activities in the face of climate
the importance of precise rainfall forecasting in
change.
Bangladesh, where millions of people rely on
agriculture for their livelihood. Accurate and Rahman et al. [7] utilized statistical techniques,
timely forecasts help farmers minimize the including linear regression and Support Vector
effects of unfavorable weather on agricultural Machine (SVM) methods, to model and predict
production by helping them plan irrigation rainfall across Bangladesh. This study addresses
strategies, grow crops at the right times, and use the pressing issue of climate change and its
resources most effectively. Furthermore, accurate impact on rainfall patterns, particularly in
rainfall forecasting is essential to disaster Bangladesh.
management initiatives because it allows
policymakers to take preventative action to lessen Mannan et al. [8] analyzed historical rainfall data
the likelihood of landslides, flooding, and other and explores two prediction models for
weather-related calamities. Artificial intelligence northeastern Bangladesh. The long-range model
technologies have a high chance of succeeding in shows promise for predicting total monsoon
these fields [3]. rainfall and individual rainfall categories in
Sylhet. The short-range model offers a way to
[Link] REVIEWS: estimate area-average and maximum rainfall,
along with rainfall amounts for Sylhet
This section highlights a few noteworthy
specifically.
studies that other authors have conducted that
are relevant to our research question. Zaman et al. [9] examined rainfall patterns in
Bangladesh over the past 30 years, focusing on
Ria et al. [3] developed a rain prediction model changes in intensity and variability across
using Machine Learning. Used total five models different seasons. Using data from 34
for the study. Each model has trained with eight meteorological stations spread throughout the
input features then it has validated the rainfall country and machine learning algorithms
predictions. But Random Forest predicts the proposed rain prediction model.
dataset most accurately so that it is the best for
this study. Hasan et al. [10]developed the most accurate
rainfall prediction model by utilizing machine
Islam et al. [4] analyzed the performance of six learning and feature selection techniques from
machine learning algorithms for predict rain, several machine learning algorithms such as
regression LGB with SelectKBest feature Naive Bayes (NB), Decision Tree, Support
selection had the best performance on the test set Vector Machine (SVM), Random Forest, and
with R2-score of [Link] set collected from Logistic Regression. Among them the Artificial
BRRI’s public climate database(5 or 6years). Neural Network (ANN) achieves a maximum
accuracy of 90%.
Paul et al. [5] developed a model for the rainfall
prediction of 34 metrological stations in Azmain et al. [11]developed a machine learning
Bangladesh and discovered that the annual approach utilizing to enhance accuracy and
average rainfall trend was declining at a rate of precision of weather forecasting. Trained
0.023 mm per year. The effects of rainfall Random Forest Regression model using weather
data of Bangladesh from 1901 to 2015 and able nation, recording different parameters like
to achieve a 91% accuracy. precipitation, temperature, mugginess, wind
speed, and barometrical weight. They frequently
Mahmud et al. [12]developed several ML give information online or through coordinate
models, namely linear regression, linear demands.
regression with parameter penalty, cross-
validation with linear regression, linear Data Preprocessing: Preparing the Data for
regression with principal component analysis Machine Learning Once the data is collected, it
(PCA), support vector regression (SVR) without needs to be transformed into a format suitable for
PCA, support vector regression (SVR) with PCA, machine learning algorithms. This crucial step,
and artificial neural networks (ANN) and trained known as data preprocessing, involves several
those models with climate data from Colorado, a key processes:
significant American metropolis, for the years
2015 to 2018. They proposed artificial neural
network achieves the best accuracy.
TABLE I. RELATED WORK TABLE..
SL. Name Year Algorithm Accuracy Data
No. Size
1. Standardization Of Rainfall 2021 Random Forest. 87.68 2,391
Prediction In Bangladesh
Using Machine Learning
Approach.
2. Evaluation of Machine 2020 LGB 73.3 4954
Learning Methods for
Predicting Rainfall in
Bangladesh.

3. Machine Learning based 2022 Random Forest, 17.40 (198

rainfall forecasting in Linear 9-
different season in Regression, 2018
Bangladesh. SVM. )

4. A Machine Learning Regression, 65.6(Ran 1948

Approach to Analyze and Random Forest, dom to
Predict Rainfall in Different Decision Forest) 2014.
Regions of Bangladesh. Tree.

5. Rainfall prediction in the 2022 Linear 1989 Fig. [Link] creation procedure.
Southeastern region of Regression, -
Bangladesh using the Linear SVM. 2019
Regression Method.
Data Cleaning:
6. Rainfall prediction over
northeastern part of
2015 WRF. 1956
- Missing Values: As mentioned earlier, identify
2014
Bangladesh during monsoon
season. and address missing data points. Techniques like
mean/median imputation or deletion (if minimal)
can be used.
III. RESEARCH METHODOLOGY Outlier Detection and Handling: Outliers can
A. Dataset skew the model's learning. Use statistical analysis
or outlier detection algorithms to identify them.
The dataset is the most significant and crucial Strategies include capping outliers to a
aspect of our research study. This segment will go reasonable range or removal (if justified).
through the information collection and Inconsistent Formats: Ensure consistency in units
arrangement prepare, counting the strategy and (e.g., temperature in Celsius or Fahrenheit) and
approaches utilized in this consider. data types (e.g., convert dates to a standard
format).
Data Sources:
TABLE II. DATASET ATTRIBUTES IN DETAILS
Bangladesh Meteorological Department Attribute Description Type

(BMD) [13]: The essential source for verifiable Date The date of the observation String

climate information in Bangladesh. The BMD Temp9am The temperature at 9 am in Celsius Numerical

keeps up a organize of climate stations over the Temp3pm The temperature at 3 pm in Celsius Numerical
MinTemp The minimum temperature in Celsius of the observation day. Numerical

MaxTemp The maximum temperature in Celsius of the observation day. Numerical

Rainfall The amount of rainfall in millimeters of the observation day. Numerical

RainToday Whether it rained today or not (Yes/No) Categorical

Evaporation The amount of evaporation in millimeters of the day. Numerical

Sunshine The number of hours of bright sunshine of the observation day. Numerical

WindGustDir The direction of the strongest wind gust of the observation day. Categorical

WindGustSpeed The speed (km/h) of the strongest wind gust of the day. Numerical

WindDir9am The direction of the wind at 9 am of the observation day. Categorical

WindDir3pm The direction of the wind at 3 pm of the observation day. Categorical

WindSpeed9am The wind speed (km/h) at 9 am of the observation day. Numerical

WindSpeed3pm The wind speed (km/h) at 3 pm of the observation day. Numerical

Humidity9am The relative humidity at 9 am of the observation day. Numerical

Humidity3pm The relative humidity at 3 pm of the observation day. Numerical

Pressure9am Atmospheric pressure at 9 am of the observation day. Numerical

Pressure3pm Atmospheric pressure at 3 pm of the observation day. Numerical

Cloud9am Fraction of sky covered by cloud at 9 am (0 to 8) Numerical

Cloud3pm Fraction of sky covered by cloud at 3 pm (0 to 8) Numerical

In our data set there is total 21 columns and 3272

rows. Those 21 columns represent 21 attributes
but for our research we use only 13 attributes. Fig. 3. Correlation matrix.
B. Classification algorithms
Logistic regression: A statistical technique for
binary classification that estimates the likelihood
that an instance will belong to a specific class is
called logistic regression. It converts the output
of linear regression into a probability between 0
and 1 using the logistic function. In order to
optimize the probability of detecting real results,
model coefficients are estimated. The two classes
are divided by a decision boundary, and
predictions are based on whether the likelihood is
higher than a predetermined threshold. It is
extensively utilized for activities like illness
diagnosis and customer churn prediction in a
variety of industries, including marketing,
finance, and healthcare.
Fig. 2. Dataset Distribution Representation.
KNN: A straightforward yet effective supervised
Figure 2 shows distribution representation of the learning technique for classification and
data for of the 9 attributes(13). And the regression problems is K-Nearest Neighbors
Correlation matrix for our dataset is shown in (KNN). It functions by locating the K data points
Figure 3. A correlation matrix is a table that in the feature space that are closest to a particular
shows the correlation coefficients forvarious query point, then classifying or regressing
variables. according to the average of these neighbors or the
majority vote. Because KNN is instance-based
and non-parametric, it depends only on the local
structure of the data and makes no assumptions
about the underlying data distribution. Although
it is simple to comprehend and apply, making it frequently utilized for activities like fraud
appropriate for a variety of uses, it can be detection, disease diagnosis, and gene expression
computationally costly for large datasets and is analysis in a variety of sectors, including finance,
dependent on the choice of K and the distance healthcare, and bioinformatics.
metric.
Ensemble classifiers: Ensemble classifiers
Decision trees: A flexible supervised learning enhance performance by merging predictions
technique for both regression and classification from various base models. Techniques including
applications is the decision tree. It creates a stacking, boosting, and bagging are frequently
decision tree by iteratively dividing the data into employed. They use the combined knowledge of
subsets according to the most important attribute various models to improve precision and
at each stage. The goal of the splitting procedure resilience. Many fields, such as natural language
is to increase the target variable's homogeneity processing and computer vision, use ensemble
within the subsets. Decision trees can handle both approaches extensively. When compared to
numerical and categorical data and are easily individual models, they provide better
interpretable and visualized. They can be generalization and accuracy.
sensitive to slight changes in the data, though, and
they are prone to overfitting, particularly with
deep trees. Pruning and ensemble techniques IV. RESULT ANALYSIS AND DISCUSSION
(like Random Forests) are frequently employed to
lessen these problems. This section provides a comparative analysis of
the supervised machine learning algorithms used
to predict rainfall probability depending on daily
Support Vector Machines (SVM): Strong
climate data. We use six machine learning
supervised learning models for classification and
regression applications include Support Vector algorithms for predict rainfall. Among them
Machines (SVM). SVM operates by determining Ensemble classification approach scored highest
the best hyperplane to divide data into distinct in rain forecast accuracy at 83.66%. The accuracy
classes while optimizing the margin—the results for each evaluated algorithm are broken
distance between the nearest data points, or out as follows:
support vectors—between the hyperplane and the
data. With the use of various kernels, including TABLE II. Accuracy Scores of Classification
linear, polynomial, and radial basis function Algorithms.
(RBF) kernels, SVM is capable of handling both SL. Algorithm F1 Accuracy
linear and non-linear classification tasks. SVM NO. Score Score
works well in high-dimensional spaces and is 1. Logistic 62.02% 82.35%
resistant to overfitting; nonetheless, it is not a regression
direct source of probability estimates and can be 2. KNN 82.52% 80%
sensitive to the choice of kernel and parameters. 3. Decision 56.54% 76.5%
trees
Random Forest : A flexible ensemble learning 4. SVM 61.86% 82.88%
technique for classification and regression 5. Random 62.15% 83.4%
applications is called Random Forest. During Forest
training, it builds several decision trees and 6. Ensemble 62.52% 83.66%
outputs the mean (for regression) or mode (for classifiers
classification) of each tree. To minimize
overfitting and boost robustness, each tree in the
forest is trained using a random selection of The findings indicate that, in comparison to
characteristics and training data. Random Forest other algorithms, KNN, Random Forest, and
offers feature importance estimations and Ensemble Classifiers performed substantially
manages high-dimensional data with ease. It's better in terms of both F1 score and accuracy
score. Of all the algorithms, KNN obtained
the highest F1 score of 82.52%. Random In conclusion, this study highlights how
Forest and Ensemble Classifiers came in machine learning methods—especially
second and third, respectively, with F1 ensemble classifiers—have the potential to
ratings of 62.15% and 62.52%. Additionally, transform Bangladesh's rainfall forecasting
Ensemble Classifiers showed the greatest and strengthen the country's agricultural and
accuracy score of 83.66%, proving that they economic sectors. To properly utilize
were capable of correctly classifying the machine learning's advantages in tackling the
dataset. These results point to these problems brought on by climate variability
algorithms' promise for the particular task and change, more study and advancement in
under investigation in this study. this area are essential.

References

[1] S. M. M. K. M. M. B. Mohammad Anisur

Rahman, "Prediction and Trends of
Rainfall Variability over Bangladesh,"
Science Journal of Applied Mathematics
and Statistics, p. 54, 2017.

[2] M. M. Md. Habibur Rahman, "On the

Fig. 5. Prediction performance visualization. Prediction of Average Monsoon Rainfall in
Bangladesh with Artificial Neural
V. CONCLUSION Network," International Journal of
This study's conclusion emphasizes the Computer Applications (0975 – 8887), vol.
noteworthy progress made possible by using 127, p. 45, 2015.
machine learning algorithms to anticipate
Bangladesh's rainfall. After conducting a [3] J. F. A. I. K. M. M. Nushrat Jahan Ria,
thorough examination of several supervised "Standardization Of Rainfall Prediction In
Bangladesh Using Machine Learning
methods, such as logistic regression, KNN,
Approach," International Conference on
decision trees, SVM, Random Forest, and
Computing and Networking Technology
ensemble classifiers, it is clear that the latter,
(ICCNT), p. 1, 2021.
with an accuracy score of 83.66%, hold great
potential for improving the accuracy of [4] R. I. A. J. a. S. M. Ferdous Zeaul Islam,
rainfall forecasts. "Evaluation of Machine Learning Methods
for Predicting Rainfall in Bangladesh,"
For the millions of people in Bangladesh who 2022 IEEE 2nd Conference on Information
depend on agriculture for a living, accurate Technology and Data Science (CITDS),
rainfall forecasts are essential for disaster 2022.
relief, agricultural planning, and the
distribution of water resources. Machine [5] M. H. R. M. H. S. Shuva Mai Paul,
learning algorithms help farmers mitigate the "Machine Learning based rainfall
effects of unfavorable weather conditions and forecasting in different season in
empower policymakers to take preventative Bangladesh.," 2022 International
action against weather-related calamities by Conference on Recent Progresses in
delivering accurate and timely forecasts.
 Science, Engineering and Technology Prediction: Unveiling Insights and
(ICRPSET), 2022. Forecasting for Improved Preparedness,"
vol. 11, pp. 132196-132222, 2023.
[6] A. N. M. M. &. B. S. A. Rahee, "machine
learning approach to analyze and predict [11] A. T. M. -U. -S. C. M. S. C. a. M. H.-E.-H.
rainfall in different regions of Bangladesh," M. A. Azmain, "Prediction of Rainfall in
2021. Bangladesh: A Case Study of the Machine
Learning," 022 IEEE 7th International
[7] M. M. H. S. a. A. A. R. Md. Habibur conference for Convergence in Technology
Rahman, "Rainfall prediction in the (I2CT), Mumbai, India, pp. 1-5, 2022.
Southeastern region of Bangladesh using
the Linear Regression Method," in [12] H. M. M. H. B. M. O. R. a. K. W. M. T.
Proceedings of the 5 th International Mahmud, "Machine Learning-based
Conference on Industrial & Mechanical Rainfall Prediction from Weather Data: A
Engineering and Operations, Dhaka, 2022. Comparative Analysis," 2023 International
Conference on Next-Generation
[8] M. A. C. M. A. M. K. S. A. S. &. M. S. J. Computing, IoT and Machine Learning
Mannan, "Rainfall prediction over (NCIM), Gazipur, Bangladesh2023
northeastern part of Bangladesh during International Conference on Next-
monsoon season.," A Scientific Journal of Generation Computing, IoT and Machine
Meteorology and Geo-Physics, Bangladesh Learning (NCIM), pp. 01-06, 2023.
Meteorological Department,, vol. 01, pp.
14-25, 2015. [13] B. Government, "Bangladesh Open Data,"
[Online]. Available:
[9] Y. Zaman, "Machine learning model on [Link]
rainfall-a predicted approach for condition. [Accessed 19 March 2024].
Bangladesh," Doctoral dissertation, United
International University, 2018.

[10] M. M. Hassan, M. A. T. Rony, M. A. R.

Khan, M. M. Hassan, F. Yasmin, A. Nag,
T. H. Zarin, A. K. B. Alshathri and W. El-
Shafai, "Machine Learning-Based Rainfall