A Machine Learning Model for Estimating Daily Rainfall in

Mediterranean Climate

Ali Karah Bash

, Amin Gharehbaghi

and Shahaboddin Daneshvar

Department of Electrical and Electronic Engineering, Faculty of Engineering, Hasan Kalyoncu University,

27110 Şahinbey, Gaziantep, Turkey

Department of Civil Engineering, Faculty of Engineering, Hasan Kalyoncu University,

27110, Şahinbey, Gaziantep, Turkey

Department of Computer Engineering, Faculty of Engineering, Hasan Kalyoncu University,

27110 Şahinbey, Gaziantep, Turkey

Keywords: Rainfall, Prediction, CGSVM Model, RBFNN Model, Türkiye.

Abstract: Rainfall estimation remains a critical yet complex task, especially in Mediterranean regions where climatic

variability poses significant modeling challenges. This study utilizes two regression approaches—Coarse

Gaussian Support Vector Machine (CGSVM) and Radial Basis Function Neural Network (RBFNN)—to

predict daily rainfall over Bozcaada station, Türkiye. The models were trained and evaluated using standard

regression performance metrics to investigate their predictive ability under Mediterranean climate conditions.

Both models showed promising results in capturing the overall structure of rainfall variation. The RBFNN

displays slightly greater stability across low-to-moderate precipitation ranges. However, neither model fully

captured the intensity of extreme rainfall events, reflecting a common limitation in data-driven rainfall

modeling. Quantitative assessments using RMSE, MAE, NSE, SI, and R² further highlighted the close

performance of both methods, with RBFNN offering marginally improved accuracy. The findings suggest

that CGSVM and RBFNN can provide useful estimations in operational contexts, though additional

enhancements are needed. This work contributes to the growing literature on machine learning applications

in hydrometeorological forecasting and highlights the need for adaptable models suited to the specific

complexities of Mediterranean climates.

1 INTRODUCTION

Freshwater resources are becoming increasingly

constrained and vulnerable as a result of

anthropogenic pollution, population growth, and

natural climate variability (Gharehbaghi and Kaya

2022; Rajput et al. ,2023, Gharehbaghi et al. 2024).

Undoubtedly, rainfall, as a primary component of

precipitation, is one of the main sources of freshwater

and constitutes a climatic event that significantly

impacts human life. Accurate rainfall forecasting can

lead to optimal planning and management in different

fields, such as agriculture, hydroelectric power

generation, water supply, and disaster prevention

(Ananth (2020); Abdel Azeem and Dev (2024);

https://orcid.org/0000-0002-6513-9180

https://orcid.org/0000-0002-2898-3681

https://orcid.org/0000-0002-0917-7254

Karah Bash et al. (2025)). Despite the significance of

accurate precipitation forecasting, it is still

challenging due to the fast-changing, uncertainty and

complexity of the atmospheric processes that affect

precipitation.

In recent decades, machine learning (ML)

techniques have provided a significant advantage in

the prediction process. “ML …can be applied to

rainfall prediction by using historical data of

meteorological variables and learning patterns or

relationships that can be used to forecast future

rainfall. Abdel Azeem and Dev (2024).

A short literature review focusing on recent

advance in this field is provided below:

Karah Bash, A., Gharehbaghi, A. and Daneshvar, S.

A Machine Learning Model for Estimating Daily Rainfall in Mediterranean Climate.

DOI: 10.5220/0014375200004848

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 2nd International Conference on Advances in Electrical, Electronics, Energy, and Computer Sciences (ICEEECS 2025), pages 295-299

ISBN: 978-989-758-783-2

295

Liyew and Melese (2021) used three ML

techniques (viz., multivariate linear regression, RF,

and Extreme Gradient Boost) to identify the relevant

atmospheric features that cause rainfall and predict

the intensity of daily rainfall. They used climate data

measured in Ethiopia and finally observed that the

Extreme Gradient Boost ML model outperformed

other models. Ojo and Ogunjo (2022) employed two

multivariate polynomial regressions (MPR) and

twelve ML models (e.g., support vector machine

(SVM) and adaptive neuro-fuzzy inference systems

(ANFIS)) for Nigeria. To this end, they used 31-year

data. The results illustrated that the adaptive ANFIS

model’s algorithms outscored the MPR, ANN, and

SVM models in the ten months of the year. Baig et

al., 2024, investigated the potential of various ML and

ensemble models, including XGBoost, Long Short-

Term Memory (LSTM), Random Forest (RF),

Gradient Boost (GB), Support Vector Machine

(SVM), Multilayer Perceptron (MLP), Linear

Regression (LR), and ensemble methods for monthly

rainfall prediction in hyperarid environments. They

reported that although initially using limited input

parameters, the models used could not provide

reliable outputs, but after adding meteorological

parameters such as wind speed, temperature,

humidity, and evapotranspiration, all models,

especially XGB and LSTM, showed significant

improvements in results. Farooq et al., 2024, utilized

2 ML models (i.e., RF and LSTM) to examine how

multiple climate indices simultaneously influence

wet-period rainfall patterns at two Northern Territory

(NT) stations in Australia. They announced that

large-scale climate factors such as the Madden Julian

Oscillation and lagged Indian Ocean Dipole

significantly influence wet-period rainfall predictions

of the NT. Moreover, the LSTM model provided

more accurate outcomes than the RF model. Mesta et

al. (2024) assessed the efficiency of ensemble

analysis for south and southwestern Türkiye. They

applied three ensemble methodologies: simple

average of the models, multiple linear regression for

super ensemble, and artificial neural networks

(ANN). The outcomes revealed that ensembled time

series performed better than individual regional

climate models.

2 MATERIALS AND METHODS

2.1 Description of the Study Area

Bozcaada is an island in the Çanakkale province of

Türkiye with a surface area of 40 km2. In the region

where this station is located, summers are warm, dry,

and clear; winters are long, cold, rainy, and partly

cloudy, and the weather is windy all year round

(Köppen classification: Csa). The temperature varies

typically between 5°C and 30°C throughout the year.

In this study, average temperature (Tmean), relative

humidity (RH), maximum wind speed (Umax), and

wind direction (Udir) were utilized as input

parameters, while rainfall was selected as the target

variable. The location of the Bozcaada station on the

map of Türkiye is depicted in Figure 1. Furthermore,

the statistical data for the Bozcaada station covering

the period from 2008 to 2019, along with its

geographical coordinates, are presented in Tables 1

and 2.

Figure 1: Location of the Bozcaada station in Türkiye.

Table 1: Statistical values of the Bozcaada station from 2008 to 2019.

Tmean RH Umax Udi

Rainfall

Mean 16.48675565 74.11213 11.73742 196.5554 1.497582

Standar

Erro

0.095480415 0.136026 0.077683 2.068534 0.102267

Median 16.9 74 11.3 180 0

Mode 22.8 73.8 13.4 360 0

Standar

Deviation 6.321207299 9.005518 5.142971 136.9457 6.770497

Sam

le Variance 39.95766172 81.09935 26.45015 18754.13 45.83962

Kurtosis -0.681687415 0.009872 3.424785 -1.62117 410.6713

Skewness -0.354301381 0.042623 1.145221 -0.049 14.90925

ICEEECS 2025 - International Conference on Advances in Electrical, Electronics, Energy, and Computer Sciences

296

Table 2: Geographical coordinates of the meteorological

station.

Station Latitude (N) Longitude(E) Elavation(M)

Bozcaada 39.8326 26.0728 30

2.2 Regression Methods

Coarse Gaussian Support Vector Machine

(CGSVM). Support Vector Machine (SVM) is a

supervised learning algorithm originally designed for

classification and regression tasks through a method

known as Support Vector Regression (SVR). In a

regression task, SVM seeks to find a function that

approximates the target variable within a specified

margin of tolerance ( 𝜀), while minimizing model

complexity. The Coarse Gaussian version refers to

using a Gaussian Kernel (GK) with a large kernel

scale, leading to smoother decision boundaries that

generalize well in cases with relatively low noise. The

Gaussian kernel function is given by:

𝐾

(

𝑥,𝑥



)

=exp −















 (1)

where 𝑥 and 𝑥



are feature vectors, and 𝜎



is the

kernel scale (larger in coarse SVMs), which controls

the spread of the Gaussian function.

Radial Basis Function Neural Network (RBFNN).

The Radial Basis Function Neural Network (RBFNN)

is a form of Artificial Neural Network (ANN) that

uses radial basis functions as activation functions.

The architecture of this method comprises three

layers:

An input layer: This layer is connected to a hidden

layer of GK neurons to provide a linear output.

The non-linear mapping of inputs: higher

dimensional space in the hidden layer of the network

takes place where summation takes place.

The RBFNN output can be mathematically expressed

as:

𝑓(𝑥) =

∑





𝑤



⋅𝜙

(‖

𝑥−𝑐



‖)

‖ (2)

where 𝑓(𝑥 is the predicted output, 𝑤



is the output

weight, 𝜙 is the radial basis function (commonly

Gaussian), 𝑐



is the center of the RBF unit, and 𝑁 is

the number of hidden neurons.

2.3 Performance Indices

Five standard evaluation metrics, including Root

Mean Square Error (RMSE), Mean Absolute Error

(MAE), coefficient of determination (R2), Scatter

Index (SI), and Nash–Sutcliffe Efficiency (NSE),

were utilized to assess the accuracy and performance

of the suggested and employed model. The

mathematical expressions for these statistical

measures are as follows:

𝑅𝑀𝑆𝐸 =







∑





𝐸𝑇



−𝐸𝑇







(3)

𝑀𝐴𝐸 =





∑





𝐸𝑇



−𝐸𝑇



 (4)

𝑅



=1−

∑

(







)



∑

(







)



(5)

𝑆𝐼 =







∑





𝐸𝑇



−𝐸𝑇







/𝐸𝑇





× 100%

(6)

𝑁𝑆𝐸 = 1 −

∑



















∑





















(7)

where the subscripts "pre" and "obs" denote

predictions and observations, respectively, and the

superscript n signifies the total number of data points.

3 RESULTS AND DISCUSSION

In this research, the outcomes of the proposed two

regression models, CGSVM and RBFNN, are

evaluated. The authors investigated the performance

and effectiveness of these models by comparing their

prediction accuracy and R2 values, which they

compute based on the prediction features from the

initial training phase.

3.1 Prediction Outcomes of CGSVM

and RBFNN Models

As shown in Figure 2, actual daily rainfall values

compare with the estimated values using the CGSVM

and RBFNN approaches. The Figure Analysis shows

that both models are usually able to reproduce the

overall trends in data, especially during periods of

low-to-moderate precipitation. RBFNN seems to

produce more stable predictions than the other two.

This is especially true for sections with denser and

less variable data. However, both models appear to be

limited in simulating pronounced spike-type rainfall.

These sudden peaks that coincide with high-intensity

rain are only partly traced by the models. This is often

the case with data-driven rainfall estimation.

Generally, there is a reasonable agreement between

the predictions and the actual measurements using

both models, indicating the potential applicability of

both models in Mediterranean climates.

A Machine Learning Model for Estimating Daily Rainfall in Mediterranean Climate

297

Figure 2: Comparison of actual and predicted daily rainfall values using CGSVM and RBFNN models

3.2 Metric Assessment Outcomes

Figure 3 presents a spider chart to compare the

performance of CGSVM and RBFNN by using five

evaluation metrics: RMSE, MAE, R2, NSE, and SI.

The Figure reveals that both models perform similarly

across most metrics, yet subtle distinctions can be

observed. The RBFNN model slightly outperforms

CGSVM in terms of RMSE and MAE, indicating a

marginally lower average error and tighter overall fit.

This suggests that the RBFNN's architecture may be

better suited for capturing the nonlinear behavior of

rainfall data. In contrast, CGSVM shows a slight

advantage in the SI and R2 metrics, implying slightly

Figure 3: Comparative radar chart of evaluation metrics for

CGSVM and RBFNN models.

better variance explanation and normalized error

performance. However, the difference between the

two models is not substantial across any single metric,

suggesting that their overall regression capabilities are

comparably effective. The NSE values for both

models remain moderate, reinforcing the earlier

observation that while the models are proficient in

capturing general trends, their ability to predict peak

rainfall events remains limited.

4 CONCLUSION

This study explored the potential of two classical

regression models—Coarse Gaussian Support Vector

Machine (CGSVM) and Radial Basis Function

Neural Network (RBFNN)—for estimating daily

rainfall in a Mediterranean climate, using data from

Bozcaada station, Türkiye. Both models were

evaluated not only in terms of visual agreement with

observed rainfall patterns but also through a set of

standard performance metrics, including RMSE,

MAE, NSE, SI, and R2. The findings indicate that

while both CGSVM and RBFNN were reasonably

effective in capturing the overall structure of daily

rainfall, RBFNN demonstrated slightly more

consistent accuracy, particularly for moderate rainfall

events. However, both models exhibited limitations

in predicting high-intensity rainfall, which is often

sparse and highly irregular. This underlines an

ongoing challenge in precipitation modeling—

0.2

0.4

0.6

0.8

RMSE

MAE

R2SI

NSE

Metric Evaluation Result

CGSVM

RBFNN

ICEEECS 2025 - International Conference on Advances in Electrical, Electronics, Energy, and Computer Sciences

298

achieving a balance between general pattern

recognition and responsiveness to extreme values.

Although the RBFNN had a slight edge in error

reduction, the overall difference between the two

approaches was relatively small, suggesting that both

can serve as viable tools for rainfall estimation in

similar climatic settings. Future studies should

consider integrating ensemble techniques or multi-

source data fusion to enhance model reliability in

Mediterranean regions.

ACKNOWLEDGEMENTS

The authors would like to thank the Turkish State

Meteorological Service for providing access to the

weather station data.

REFERENCES

Abdel Azeem, M., Dev, S., (2024) A performance and

interpretability assessment of machine learning models

for rainfall prediction in the Republic of Ireland,

Decision Analytics Journal 12, 100515,

https://doi.org/10.1016/j.dajour.2024.100515

Ananth, J., (2020) MapReduce and optimized deep network

for rainfall prediction in agriculture, Comput. J., 63 (6),

900–912.

Baig, F., Ali, L., Faiz, M.A., Chen, H., Sherif, M. (2024)

How accurate are the machine learning models in

improving monthly rainfall prediction in hyper arid

environment?, Journal of Hydrology 633, 131040,

https://doi.org/10.1016/j.jhydrol.2024.131040

Farooq, R., Imteaz, M.A., Shangguan, D., Hlavčová, K.,

(2024) Machine learning algorithms to forecast wet-

period rainfall using climate indices in Northern

Territory of Australia, Science Talks 12, 100397,

https://doi.org/10.1016/ j.sctalk.2024.100397

Gharehbaghi, A., Afaridegan, E., Kaya, B., Adhami, M.

(2024) Calibration and evaluation of various reference

evapotranspiration estimation methods in a humid

subtropical climate: A case study in Samsun Province,

Türkiye, Physics and Chemistry of the Earth, Parts

A/B/C, 136, 103734,

https://doi.org/10.1016/j.pce.2024.103734.

Gharehbaghi, A., Kaya, B., (2022) Calibration and

evaluation of six popular evapotranspiration formula

based on the Penman-Monteith model for continental

climate in Turkey. Phys. Chem. Earth, Parts A/B/C 127,

103190. https://doi.org/ 10.1016/j.pce.2022.103190.

Karah Bash AA, and Khail, AA (2025) Advanced multi-

layer deep learning model for accurate estimation of

heat transfer and flow designing parameters across

diverse dataset configurations. Engineering

Applications of Artificial Intelligence 151: 110723.

https://doi.org/10.1016/j.engappai.2025.110723

Liyew C.M. and Melese, H.A. (2021) Machine learning

techniques to predict daily rainfall amount, Journal of

Big Data 8:153, https://doi.org/10.1186/s40537-021-

00545-4

Mesta, B., Akgun, B., Kentel, E., (2024) Improving

precipitation estimates for Turkey with multimodel

ensemble: a comparison of nonlinear artificial neural

network method with linear methods, Neural

Computing and Applications, 36:10219–10238,

https://doi.org/10.1007/s00521-024-09598-x

Ojo, O.O., Ogunjo, S.T., (2022) Machine learning models

for prediction of rainfall over Nigeria, Scientific

African, 16, e01246,

https://doi.org/10.1016/j.sciaf.2022.e01246

Rajput, J., Singh, M., Lal, K., Khanna, M., Sarangi, A.,

Mukherjee, J., Singh, Sh, 2023. Performance evaluation

of soft computing techniques for forecasting daily

reference evapotranspiration. Journal of Water and

Climate Change 14 (1), 350. https://doi.

org/10.2166/wcc.2022.385.

A Machine Learning Model for Estimating Daily Rainfall in Mediterranean Climate

299