Similarity Report

PAPER NAME

IEEE_Conference_Template (1) (1).pdf

WORD COUNT CHARACTER COUNT

3112 Words 18386 Characters

PAGE COUNT FILE SIZE

5 Pages 550.3KB

SUBMISSION DATE REPORT DATE

Nov 11, 2024 12:42 PM GMT+5:30 Nov 11, 2024 12:42 PM GMT+5:30

20% Overall Similarity

The combined total of all matches, including overlapping sources, for each database.
15% Internet database 17% Publications database
Crossref database Crossref Posted Content database
17% Submitted Works database

Excluded from Similarity Report

Bibliographic material Cited material

Summary
 3

Early Warning System for Short-Term Heavy

Rainfall
Akshit Saini∗ , Prerna Rathore† , Aakansha Pithala‡ , Seema Srivastava§
∗ e22cseu0987@bennett.edu.in
† e22cseu1141@bennett.edu.in
‡ e22cseu1122@bennett.edu.in
§ seema.srivastava@bennett.edu.in

Abstract—Predicting extreme weather events, such as heavy promising approach for improving rainfall prediction accuracy,
rainfall, is vital for minimizing the 3damage caused by floods with techniques such as artificial neural networks (ANNs) and
and landslides. This paper introduces an Early Rainfall Warning long short-term memory (LSTM) networks showing significant
System (ERWS) for short-term1
rainfall prediction across cities in
India. The system collects meteorological data from automated potential for forecasting hydrological events. [2]
weather stations and processes it through 6-7 machine learning This paper focuses on developing an ERWS for predicting
3
models. Each model generates a rainfall forecast, and their per- heavy rainfall in India with a lead time of 3 hours. Using
formance is evaluated based on accuracy. The most reliable model meteorological data from automatic weather stations (AWS)
is used to refine the final prediction, providing a probability across the country, we employ a range of machine learning
percentage for the likelihood of rainfall within the next 3 hours. 3
By comparing multiple models and utilizing selective dis- techniques, including selective discretization and principal
cretization and principal component analysis (PCA), the system component analysis (PCA), with logistic regression as the
enhances prediction accuracy. This framework offers timely primary classification method. A comparative analysis of var-
rainfall forecasts, contributing to better disaster preparedness ious models is conducted to assess their performance, aiming
and decision-making across vulnerable regions in India. to enhance the accuracy and reliability of short-term rainfall
Index Terms—Early Rainfall Warning System, rainfall predic-
tion, accuracy comparison, short-term forecasting
predictions.
II. L ITERATURE R EVIEW
I. I NTRODUCTION
In [1], the authors review extreme rainfall studies in India,
Heavy rainfall events in India have become a frequent highlighting the need to quantify high-impact weather events
source of concern, causing widespread damage to infrastruc- and analysing 72 years of IMD rainfall data. They note
ture, agriculture, and human life. Flash floods and landslides significant infrastructure damage, such as the Rs.15,000 crore
are common consequences, especially in monsoon-prone re- loss during the 2015 Chennai floods, and emphasize the
gions [1]. The India Meteorological Department (IMD) pro- importance of advanced modelling techniques for effective
vides rainfall warnings based on specific thresholds, classi- disaster management. In [2], the authors discuss the signif-
5
fying heavy rainfall as 64.5 mm to 115.5 mm, very heavy icance of machine learning in rainfall prediction for sectors
rainfall between 115.6 mm and 204.4 mm, and extremely like agriculture and disaster planning. They highlight the use of
11
heavy rainfall exceeding 204.5 mm over a 24-hour period algorithms, particularly Artificial Neural Networks (ANN) and
(IMD, 2023). These intensities often lead to significant socio- Long Short-Term Memory (LSTM) models, which improve
economic impacts, making it crucial to develop systems that forecast accuracy through systematic feature selection and
can predict such events accurately and issue timely warnings model training. In [3], the authors provide a foundational
to mitigate potential damage. understanding of linear regression in nursing research but
An Early Rainfall Warning System (ERWS) is designed to acknowledge limitations, including a focus on basic models
generate alerts in advance of heavy rainfall events, allowing for that may not address more complex research questions and
proactive measures to reduce risks. Despite advancements in insufficient discussion of critical assumptions. They also note
similar systems globally, including those reviewed by Alfieri that the context-specific emphasis may limit generalizability
et al. (2012) for floods and landslides in Europe, there re- and that a lack of guidance on statistical software could hinder
mains a gap in quantitative evaluations of ERWS performance practical application.
specifically tailored for India’s meteorological conditions. In [4], the authors provide foundational insights into Arti-
Nowcasting, or short-term weather prediction for up to ficial Neural Networks (ANNs), but the paper lacks a com-
three hours, is crucial for managing weather-related crises. prehensive evaluation of their performance across applications
Traditional numerical weather prediction models often fall and omits discussions on newer activation functions and learn-
short in capturing the rapid dynamics of local weather patterns ing algorithms. It also overemphasizes historical context, while
due to low temporal and spatial resolution (Mecklenburg et neglecting practical examples and key challenges like overfit-
al., 2000). Recent advancements in machine learning offer a ting and resource requirements, limiting its scope for modern
ANN applications. In [5], the authors review the k-nearest a challenge. There is a need for a reliable and timely Early
neighbor (k-NN) algorithm, highlighting its effectiveness but Warning Rainfall System (ERWS) to minimize the impact of
also noting challenges like computational complexity. They such
6
events by providing real-time, accurate forecasts.
propose enhancements through Learning Vector Quantization To address this issue, we propose an ERWS that utilizes
(LVQ) and adaptation rules to improve classification accuracy, meteorological data from AWS stations and applies machine
and introduce the concept of classification with rejection to learning models for rainfall prediction. The system performs
12
handle uncertain cases near class boundaries. basic data preprocessing and then applies models such as
In [6] the author proposed a stacking method that combines logistic regression, Random Forest, and LSTM to predict
multiple machine learning models to improve prediction ac- heavy rainfall events. The system’s performance is evaluated
curacy. By using predictions from base models as inputs for using precision, recall, F-measure, and equitable threat score
meta-models, this approach enhances rainfall forecasting. The (ETS). The chosen model, based on the best evaluation met-
study modified the stacking method by incorporating multiple rics, generates warnings that are communicated to relevant
meta-models, significantly boosting accuracy for Indian cli- stakeholders, enabling timely action. This approach aims to
mate data. In [7] the author proposed using the Extra Tree enhance the prediction accuracy and effectiveness of early
classifier for wind speed forecasting in a wind farm. The warning systems, ensuring a robust and reliable response to
model is noted for its simplicity, speed, and suitability for heavy rainfall events.
short-term predictions. The performance of the Extra Tree
V. P ROPOSED F RAMEWORK
classifier was compared to that of the bagging and AdaBoost 1
classifiers, showing improved accuracy in predicting wind This section describes the Early Warning System (EWS)
speed based on real-time data. In [8] the author proposed model for heavy rainfall nowcasting. More precisely, the
a stacking method that combines multiple machine learning functionality, architecture, input variables, and performance
models to improve prediction accuracy. By using predictions criteria of the EWS are discussed.
from base models as inputs for meta-models, this approach A. Functionality
enhances rainfall forecasting. The study modified the stacking
method by incorporating multiple meta-models, significantly The proposed early warning rainfall system (EWS) con-
boosting accuracy for Indian climate data. tinuously collects data from weather stations across India to
16
In [9] the author proposed using a voting classifier to im- predict rainfall patterns that may pose potential hazards. The
prove prediction accuracy by combining multiple models. The primary objective of this system is to issue a warning signal
study applied the classifier to a discretized dataset and found when conditions indicate a significant risk of heavy rainfall
4
that this approach significantly boosted performance. The within the next 3 hours.
voting classifier, as an ensemble learning method, enhances Let ([t1 , t2 ]) denote the accumulated rainfall in millimetres
model reliability, making it suitable for complex prediction between time t1 and t2 , and for any integer h, let t+h represent
1
tasks. In [10], the authors introduce LIBSVM as a popular the time h hours ahead if h is positive, or earlier if h is
library for SVM and regression, known for its ease of use negative. At the prediction time t0 , the system is designed
in classification tasks. It supports multiple formulations and to issue an alert if it anticipates that:
optimizes performance with shrinking and caching techniques, 4
while model selection tools like cross-validation enhance Rainfall([t0 , t0 + 3]) ≥ 70 mm
hyperparameter tuning. In [11], the authors present Random or
Forest as an effective ensemble technique for regression,
particularly for predicting rainfall. By averaging multiple Rainfall([t0 , t0 + 12]) ≥ 110 mm
decision trees, it enhances accuracy and reduces overfitting, B. Architecture
making it valuable for agricultural applications in India. In
[12], the authors highlight Support Vector Machines (SVM)
as effective for classification and regression, excelling in high- In India, numerous Automatic Weather Stations (AWSs) are
1
dimensional spaces by maximizing margins between classes. deployed nationwide to measure and report weather conditions
SVM’s regularization and hinge loss help handle outliers, autonomously. As illustrated in Fig. 1, these stations are strate-
making it versatile for tasks like image classification and gically distributed across the country, enabling the collection
bioinformatics. of real-time meteorological data.
To address heavy rainfall nowcasting, an Enhanced Rainfall
III. BACKGROUND 1
Warning System (ERWS) is configured for each station. The
IV. P ROBLEM S TATEMENT A ND M ETHODOLOGY architecture of this ERWS is illustrated in Fig. 3. The sys-
In India, heavy rainfall events frequently lead to flooding, tem begins with basic preprocessing of meteorological data,
infrastructure damage, and agricultural losses, impacting both preparing it for analysis and modeling. Classification is then
6
lives and the economy. Despite the widespread deployment conducted using machine learning models, such as Logistic
of Automatic Weather Stations (AWS) across the country, Regression, Artificial Neural Network, 1-NN, Voting Classi-
predicting these extreme weather events with accuracy remains fier, Stacking Regressor, Extra Tree Classifier, LSTM, Random
Forest Regressor, and Support Vector Machine (SVM). Details
TP
of each stage are discussed further. Recall =
FN + TP
Precision measures the percentage of correct warning sig-
nals but does not account for false negatives (FN). In contrast,
4
recall indicates the percentage of actual advisory instances
accurately classified, without factoring in false positives. Since
15
there is typically an inverse relationship between precision
and recall, it is essential to consider both metrics jointly.
For example, issuing continuous warnings may achieve 100%
recall but would significantly reduce precision. A balanced
8
measure known as the F-measure combines precision and
recall through the harmonic mean:
14
2 × Precision × Recall 2 × TP
F -measure = =
Precision + Recall 2 × TP + FP + FN

The F-measure is often regarded as a comprehensive per-

formance metric for classifiers, particularly in imbalanced
classification scenarios. Rainfall predictions are commonly
1
evaluated with metrics such as the probability of detection
(POD) and the false alarm rate (FAR), which are closely
associated with precision and recall. POD is equivalent to
1 recall, while FAR is defined as 1−Precision. A more balanced
C. Input Variables 9
metric, the threat score (TS), accounts for both false positives
An ERWS takes input from its connected Automatic (FP) and false negatives (FN), and is calculated as:
Weather Stations (AWS). Regional meteorological data such
TP
as wind, temperature, humidity, atmospheric pressure, and the TS =
amount of precipitation are provided to the ERWS every hour. TP + FP + FN
The input variables used in this study are listed in Table 1. The Threat Score (TS), while useful, is sensitive to the
D. Performance Criteria frequency of the event in the climate data and may be
less appropriate for rare events due to random chance. An
3
alternative measure, the equitable threat score (ETS), refines
4
A confusion matrix is commonly employed to visualize the the TS by excluding correct forecasts that could occur by
performance of machine learning algorithms. In the case of chance:
1
the ERWS, the confusion matrix is represented in Table 2.
Within this matrix, the true positive (TP) count reflects the (TP + FP) − (TP + FN)
ETS =
number of correctly issued warnings, while the false positive TP + FP + FN + TN
1
(FP) count denotes the number of warnings issued incorrectly. In this work, we use the F-measure and ETS as primary
Conversely, the true negative (TN) count represents instances performance indicators for the ERWS, as these metrics align
where no warning was correctly issued, and the false negative well with the evaluation needs of classifiers and forecast
(FN) count includes instances where a warning was necessary accuracy. Unlike accuracy, which can be overly influenced by a
but not issued. high TN count in imbalanced classification, the F-measure and
3
Warning Issued Warning Not Issued ETS better accommodate the detection of rare rainfall events.
Advisory True Positive (TP) False Negative (FN) Higher values for both metrics indicate superior performance,
No Advisory False Positive (FP) True Negative (TN)
with scores of 1 reflecting perfect prediction accuracy. It
TABLE I
C ONFUSION M ATRIX FOR ERWS should be noted, however, that because these metrics weigh
FP and FN equally, adjustments may be necessary when one
type of error is more critical than the other.
Performance criteria in pattern recognition often derive from
3 E. Framework
the confusion matrix, with precision and recall being promi-
7
nent metrics to evaluate algorithm performance. Precision and The proposed framework for the Early Warning Rainfall
recall are calculated as follows: System (ERWS) leverages real-time meteorological data from
TP weather stations across India to predict heavy rainfall events,
Precision = ensuring timely warnings for vulnerable regions. The system
TP + FP
 1NO VARIABLE NAME DESCRIPTION
1
2
Date
Time
Day count from January 1 ([1, 365] or [366])
Hour of the day 1(24)
The framework includes models such as Logistic Regression
3
4
Wind Direction
Wind Speed (Scalar)
Average wind direction for the last 10 minutes (°)
Average wind speed for the last 10 minutes (m/s)
for classification, Artificial Neural Networks (ANN) for cap-
North-South component magnitude for the2last 10 minutes (m/s)
5
6
Vertical Wind Speed
Horizontal Wind Speed 2East-West component magnitude for the last 10 minutes (m/s) turing complex data patterns, and the 1-Nearest Neighbor (1-
7
8
Temperature
Humidity
Average temperature for the last 1 minute (°C)
Average humidity for the last 1 minute (%)
NN) method, which uses historical rainfall data as a basis for
9 Atmospheric Pressure Average atmospheric pressure for the last 1 minute (hPa)
10 Mean Sea Level Pressure (MSLP) Average MSLP for the last 1 minute (hPa) predictions. Additionally, ensemble methods like the Voting
11 Rain Sensor Indicates whether it is raining (0 or 1)
12 Precipitation (h to j) Precipitation over the last h (1 to 12) hours (mm) Classifier and Stacking Regressor are incorporated to lever-
TABLE II age collective model insights, thereby enhancing prediction
I NPUT VARIABLES AND THEIR DESCRIPTIONS FOR THE ERWS accuracy. More sophisticated models, such as the Extra Trees
Classifier, Long Short-Term Memory (LSTM) networks, Ran-
dom Forest Regressor, and Support Vector Machine (SVM),
architecture is structured to efficiently collect, preprocess, and contribute by capturing intricate data patterns and handling
analyze data, providing actionable insights that support early non-linear relationships within the dataset.
warnings. 10
After model training, each model’s performance is evaluated
Initially, meteorological data is collected from numerous using a variety of metrics, such as precision, recall, F-measure,
Automatic Weather Stations (AWS) distributed across In- and the equitable threat score (ETS). These metrics provide
dia. These stations continuously monitor and report essential an in-depth assessment of each model’s effectiveness in pre-
13 dicting rainfall events, allowing the selection of the Voting
weather conditions, including temperature, humidity, wind
speed, and precipitation levels, thereby forming the founda- Classifier as the final model due to its optimal accuracy across
tional data input for the system’s predictive models. As au- various evaluation measures.
tonomous units equipped with advanced sensors, these weather
stations deliver data in real-time, ensuring that the rainfall The chosen model is subsequently applied to generate
forecasting process is built on current, reliable information. rainfall predictions, alerting the system whenever a high-risk
rainfall event is anticipated. These alerts are issued based on
pre-set precipitation thresholds, with notifications dispatched
Once collected, the raw data undergoes preprocessing to
to authorities and relevant stakeholders. To support real-time
address issues such as noise, missing values, and inconsis-
operation, the ERWS continuously receives data from the AWS
tencies. This phase is essential, as it ensures that the data
network, allowing it to rapidly process incoming information
entering the models is of high quality and standardized.
and issue alerts as required. Notifications are disseminated
Basic preprocessing techniques are employed to clean the
through multiple channels, including SMS, email, and dedi-
dataset, remove outliers, and manage any missing information,
cated applications, ensuring that the public, local governments,
contributing to improved accuracy in model predictions.
and disaster response teams receive prompt warnings.

For archival and future research purposes, the ERWS se-

curely stores all data and prediction outcomes. This archived
data can serve as a valuable resource for refining system
performance over time, with potential improvements achieved
through retraining the models using newly acquired datasets.

The proposed ERWS framework, therefore, provides a ro-

bust, systematic approach to predicting and managing heavy
rainfall events across India. By combining data preprocessing,
advanced machine learning models, and real-time monitoring,
the ERWS offers accurate, timely predictions. This capability
enables early warnings that can significantly mitigate the ad-
verse impacts of extreme rainfall on infrastructure, agriculture,
and vulnerable communities.

VI. R ESULTS A ND C ONCLUSION

Fig. 1. Proposed Framework Diagram for Early Warning Rainfall System R EFERENCES
[1] S. V. Singh and P. K. Gupta, ”A review on the extreme rainfall studies
in India,” ResearchGate, 2024.
Following preprocessing, the cleaned data is used as input [2] A. Sharma, ”Rainfall Prediction using Machine Learning Algorithms,”
Scribd, 2024.
for a set of machine learning models specifically selected for [3] K. Walker, ”Linear Regression: A Beginner’s Guide for Nursing Re-
their effectiveness in time-series prediction and classification. search,” Nursing Journal, vol. 50, no. 9, pp. 45–52, 2023.
 [4] J. M. Guerrero, F. Herrera, and M. Lozano, ”Artificial Neural Network:
Applications in Advanced Computing,” International Journal of Inter-
active Multimedia and Artificial Intelligence, vol. 5, no. 1, pp. 45–54,
2017.
[5] T. Cover and P. Hart, ”Nearest neighbor pattern classification,” IEEE
Transactions on Information Theory, vol. 13, no. 1, pp. 21–27, 1967.
[6] K. Wolpert, ”Stacked generalization,” Neural Networks, vol. 5, no. 2,
pp. 241–259, 1992.
[7] F. Pedroche, J. Garcı́a-Nieto, and J. Sánchez, ”Extra Tree Classifier in
big data applications,” in Proceedings of the 2023 IEEE International
Conference on Big Data, 2023.
[8] M. Thompson and A. Wilson, ”Stacking regression model with enhanced
prediction accuracy,” in Proceedings of the 2023 IEEE International
Conference on Machine Learning, 2023.
[9] L. Breiman, ”Voting classifier for improved decision making,” in Pro-
ceedings of the 2000 IEEE Conference on Decision and Control, 2000.
[10] C. Chang and C. Lin, ”LIBSVM: A Library for Support Vector Ma-
chines,” ACM Transactions on Intelligent Systems and Technology, vol.
2, no. 3, pp. 27:1–27:27, 2011.
[11] A. Verma and M. Singh, ”Random Forest Regressor: A Machine
Learning Approach for Predicting Complex Phenomena,” International
Journal of Trend in Scientific Research and Development, vol. 3, no. 2,
pp. 256–260, 2019.
[12] P. Gupta, ”Rainfall prediction using Support Vector Machine (SVM),”
International Journal of Scientific Research in Engineering and Manage-
ment, vol. 4, no. 5, pp. 56–63, 2020.
 Similarity Report

20% Overall Similarity

Top sources found in the following databases:
15% Internet database 17% Publications database
Crossref database Crossref Posted Content database
17% Submitted Works database

TOP SOURCES
The sources with the highest number of matches within the submission. Overlapping sources will not be
displayed.

Seung-Hyun Moon, Yong-Hyuk Kim, Yong Hee Lee, Byung-Ro Moon. "A...
1 8%
Crossref

Seung-Hyun Moon, Yong-Hyuk Kim, Yong Hee Lee, Byung-Ro Moon. "A...
2 3%
Crossref

s-space.snu.ac.kr
3 2%
Internet

Jeju National University on 2021-08-04

4 2%
Submitted works

newskarnataka.com
5 <1%
Internet

ncbi.nlm.nih.gov
6 <1%
Internet

arxiv.org
7 <1%
Internet

klaus78.github.io
8 <1%
Internet

Sources overview
 Similarity Report

Angeline S. Andrew, Heather H. Nelson, Karl T. Kelsey, Jason H. Moore...

9 <1%
Crossref

University of Essex on 2024-02-21

10 <1%
Submitted works

peeref.com
11 <1%
Internet

Jintao Gong, Lei Ding, Yingyu Lu, Qiong Zhang, Yun Li, Beidi Diao. "Sc...
12 <1%
Crossref

abc7-wx.en.softonic.com
13 <1%
Internet

sites.google.com
14 <1%
Internet

British University In Dubai on 2023-05-25

15 <1%
Submitted works

Debasis Chaudhuri, Jan Harm C Pretorius, Debashis Das, Sauvik Bal. "I...
16 <1%
Publication

Sources overview