www.nature.

com/scientificreports

OPEN Predicting rainfall using machine

learning, deep learning, and time
series models across an altitudinal
gradient in the North-Western
Himalayas
Owais Ali Wani1, Syed Sheraz Mahdi1,9, Md. Yeasin2, Shamal Shasang Kumar3,
Alexandre S. Gagnon4, Faizan Danish 5, Nadhir Al-Ansari6, Salah El‑Hendawy 7&

Mohamed A. Mattar8
Predicting rainfall is a challenging and critical task due to its significant impact on society. Timely
and accurate predictions are essential for minimizing human and financial losses. The dependence
of approximately 60% of agricultural land in India on monsoon rainfall implies the crucial nature of
accurate rainfall prediction. Precise rainfall forecasts can facilitate early preparedness for disasters
associated with heavy rains, enabling the public and government to take necessary precautions. In the
North-Western Himalayas, where meteorological data are limited, the need for improved accuracy in
traditional modeling methods for rainfall forecasting is pressing. To address this, our study proposes
the application of advanced machine learning (ML) algorithms, including random forest (RF), support
vector regression (SVR), artificial neural network (ANN), and k-nearest neighbour (KNN) along with
various deep learning (DL) algorithms such as long short-term memory (LSTM), bi-directional LSTM,
deep LSTM, gated recurrent unit (GRU), and simple recurrent neural network (RNN). These advanced
techniques hold the potential to significantly improve the accuracy of rainfall prediction, offering hope
for more reliable forecasts. Additionally, time series techniques, including autoregressive integrated
moving average (ARIMA) and trigonometric, Box-Cox transform, arma errors, trend, and seasonal
components (TBATS), are proposed for predicting rainfall across the altitudinal gradients of India’s
North-Western Himalayas. This approach can potentially revolutionise how we approach rainfall
forecasting, ushering in a new era of accuracy and reliability. The effectiveness and accuracy of the
proposed algorithms were assessed using meteorological data obtained from six weather stations
at different elevations spanning from 1980 to 2021. The results indicate that DL methods exhibit
the highest accuracy in predicting rainfall, as measured by the root mean squared error (RMSE) and
mean absolute error (MAE), followed by ML algorithms and time series techniques. Among the DL
algorithms, the accuracy order was bi-directional LSTM, LSTM, RNN, deep LSTM, and GRU. For the
ML algorithms, the accuracy order was ANN, KNN, SVR, and RF. These findings suggest that altitude
significantly affects the accuracy of the models, highlighting the need for additional weather stations
in this mountainous region to enhance the precision of rainfall prediction.
1Division of Agronomy, Faculty of Agriculture Wadoora, Sher-e-Kashmir University of Agricultural Sciences &
Technology of Kashmir (SKUAST-K), Jammu and Kashmir 193201, India. 2ICAR-Indian Agricultural Statistics
Research Institute, New Delhi 110 012, India. 3Department of Agronomy (Rootcrops), Ministry of Agriculture &
Waterways (MOA & W), Suva City 679, Fiji. 4School of Biological and Environmental Sciences, Liverpool John Moores
University, Liverpool L3 3AF, UK. 5Department of Mathematics, School of Advanced Sciences, VIT-AP University,
Inavolu, Andhra Pradesh 522237, India. 6Civil, Environmental and Natural Resources Engineering, Lulea University
of Technology, 97187 Lulea, Sweden. 7Department of Plant Production, College of Food and Agricultural Sciences,
King Saud University, P.O. Box 2460, Riyadh 11451, Saudi Arabia. 8Department of Agricultural Engineering, College
of Food and Agriculture Sciences, King Saud University, P.O. Box 2460, Riyadh 11451, Saudi Arabia. 9Advanced
Centre for Rainfed Agriculture (ACRA), Dhiansar, Bari-Brahmana-181133, SKUAST-Jammu, UT-J&K, India. email:
[email protected]; [email protected]; [email protected]

Scientific Reports | (2024) 14:27876 | https://doi.org/10.1038/s41598-024-77687-x 1

www.nature.com/scientificreports/

Climate change implications on food security are significantly connected to its profound impact on agriculture.
Anticipating the conditions for the upcoming planting season becomes challenging due to the prevailing
uncertainty caused by the unpredictable nature of climate variability, often proving detrimental to agricultural
activities1. As a result, farmers and farming decision-makers heavily rely on their understanding of regional
climatic patterns when making crucial decisions about ploughing, seeding, and managing their crops. However,
traditional approaches have become less reliable in the changing climate2. Enhanced climate predictions offer
hope for improving decision-making in the agricultural sector. These advanced forecasts have the potential to
mitigate the adverse effects of factors such as a poor or delayed monsoon season and provide an opportunity
to leverage projected favourable weather conditions3. Embracing these predictions can help farmers and other
agricultural professionals navigate uncertainty more effectively, safeguarding their crops and yields while laying
the groundwork for sustainable farming practices that adapt to climate change4. Additionally, heavy precipitation
can lead to flooding, impacting infrastructure, transport networks, and human livelihood5. Therefore, it would
be advantageous to the decision-making process if the potential magnitude of rainfall over a region could be
quantified in advance. Predicting rainfall is crucial for improving agricultural output and ensuring a country’s
residents’ access to food and clean water6.
The link between climate predictions and agriculture highlights the crucial role of accurate forecasts in
shaping the future of food security. In recent decades, improving rainfall forecasting has been a focal point in
the scientific community7. To improve the accuracy of rainfall forecasts, it is essential to integrate various data
sources and utilise advanced modeling techniques. Numerical weather prediction models utilise mathematical
equations to simulate the atmosphere’s behaviour and interactions with various climatic factors. These models
incorporate historical data, real-time observations from weather stations, satellite imagery, and sophisticated
computational algorithms8. Rainfall is intricately linked to a network of climatic factors, each influencing
and being influenced by others. These factors include maximum and minimum temperatures, atmospheric
pressure, relative humidity, and wind speed9. The interconnectedness of these variables creates a complex
system that plays a crucial role in maintaining overall climate balance. Numerical weather prediction models
aim to provide more precise forecasts by examining the intricate relationships between climatic elements and
rainfall. The intricate and ever-changing nature of the atmosphere and the complex interplay of various climatic
factors have challenged the improvement of rainfall forecasting. Although there has been notable progress in
enhancing rainfall prediction techniques, several factors continue to contribute to the complexity of the task.
One significant obstacle is the inherent unpredictability of atmospheric processes10. The atmosphere is dynamic
and chaotic, with sudden shifts and unexpected interactions. This unpredictability suggests that even minor
variations in initial conditions can lead to vastly different outcomes over time11. This phenomenon presents a
substantial challenge for meteorologists and climate scientists working to create accurate models for predicting
rainfall patterns. Another crucial aspect complicating rainfall forecasting is the need for long-term historical
data to construct reliable prediction models. Historical data aids scientists in identifying rainfall patterns, trends,
and cycles. However, obtaining comprehensive and high-quality historical data is not always straightforward,
particularly in areas with limited monitoring infrastructure12.
Rainfall prediction involves complex stochastic and nonlinear behaviours, which can be addressed using
advanced techniques such as data mining, artificial intelligence (AI), ML, and DL. ML methods can reveal
hidden patterns in historical rainfall data and have been proposed as an alternative modeling approach for
nonlinear and dynamic systems5. For this reason, in recent years, ML approaches have emerged as powerful
successors to traditional data mining techniques in the domain of rainfall prediction, reflecting the growing
recognition of ML methods’ capabilities in tackling the intricate challenges of predicting precipitation
patterns13–15,16. demonstrated that ML methods are superior to traditional deterministic methods for rainfall
prediction. Examples of ML models include RF, SVR, and SVM. RF employs an ensemble of decision trees to
make predictions, enhancing accuracy and handling complex data relationships17. SVR, a form of SVM adapted
for regression tasks, effectively captures nonlinear patterns in data. SVM is a versatile model for classification
and regression, creating optimal decision boundaries through support vectors18.
DL, as a subset of ML, has also demonstrated significant potential in enhancing predictive capabilities
by utilising sophisticated neural networks inspired by the interconnected neurons of the human brain19. DL
techniques encompass various algorithms of ANNs. These networks consist of interconnected layers of nodes,
or “neurons,” each processing and transforming input data before passing it to the next layer. The depth and
intricacy of these networks enable them to capture complex patterns and relationships within extensive datasets.
This capability is particularly valuable in domains characterised by high-dimensional and nonlinear data,
such as climate science and meteorology20. DL methods are closely related to traditional ML methods, albeit
differing in architectural complexity and the hierarchy of feature extraction21. While both DL and traditional ML
strive to identify patterns in data, DL models excel in autonomously learning data representations at multiple
levels of abstraction. This implies that DL models can automatically discover intricate features within raw data
without explicit feature engineering, which often demands domain expertise and can be time-consuming22. DL
application in rainfall prediction involves training neural networks on historical climate data. These networks
are designed to identify hidden correlations, nonlinear relationships, and temporal dependencies among crucial
variables for accurate rainfall forecasts23. As the neural networks process and learn from these datasets, they
refine their internal representations, progressively enabling them to make increasingly accurate predictions. DL
methods effectively harness the interconnectedness of climatic factors in rainfall forecasting, considering diverse
variables such as temperature, humidity, wind speed, and atmospheric pressure and recognising their combined
impact on rainfall patterns24. This comprehensive approach provides an advantage over traditional methods that
may struggle to capture the intricate interactions among these variables. Furthermore, DL’s ability to process
vast amounts of data aligns well with the requirements of meteorological forecasting, where historical climate
records span many decades and comprise an array of variables. The capacity of DL models to identify subtle

Scientific Reports | (2024) 14:27876 | https://doi.org/10.1038/s41598-024-77687-x 2

www.nature.com/scientificreports/

trends, nonlinear dependencies, and intricate temporal patterns within these datasets can lead to more accurate
and reliable rainfall predictions25.
Examples of DL approaches include ANN, RNN, KNN, and GRU. ANN simulates interconnected neurons to
capture complex relationships in data, enhancing learning capabilities. RNN specialises in sequence modeling,
preserving the memory of previous inputs for tasks like language processing26. KNN is a learning algorithm that
makes predictions based on the proximity of data points in the feature space. GRU, a variant of RNN, addresses
vanishing gradient issues in RNN and improves long-range dependency capture27. Additionally, there are time
series models such as ARIMA, LSTM, trigonometric, Box-Cox transform, ARMA, and TBATS models. These
models allow data analysts to model and forecast time series data across various applications. For instance,
ARIMA combines autoregressive and moving average components with differencing for handling non-stationary
data. LSTM is an RNN that captures intricate temporal relationships in sequences28,29. Trigonometric models
leverage sinusoidal functions to capture cyclic patterns, which are ideal for data with periodic fluctuations. The
Box-Cox transform stabilises variance and enhances normality in data30. ARMA models blend autoregressive
and moving average components, while TBATS considers complex trends and seasonal patterns17. Using linear
regression as an ML approach aims to predict rainfall by establishing relationships with other atmospheric
variables31,3233. conducted a thorough comparative study to assess the effectiveness of statistical modeling and
regression techniques in predicting rainfall using environmental features. The study highlighted the superior
performance of regression techniques over statistical modeling when forecasting rainfall patterns. Additionally,
their investigation demonstrated that the RF model exhibited enhanced predictive accuracy among ML
algorithms compared to SVM and Decision Tree methods. This study contributes valuable insights into rainfall
prediction, shedding light on the advantages of specific modeling approaches. The DL LSTM model has proven
to be effective for rainfall prediction. In a study on forecasting rain in the Hyderabad region of India20,34, put
forth an enhanced LSTM model, which they compared with other models such as Holt-Winters, ARIMA,
extreme learning machine (ELM), and RNN. The results were verified using ANN to predict the monthly average
rainfall35. The findings indicate that out of the three different types of networks (layer recurrent, cascaded feed
forward back propagation, and feed forward back propagation), the feed forward back propagation network type
yielded the best results.
Previous studies have utilised a linear regression model to pinpoint crucial characteristics for predicting
rainfall, including solar radiation, detectable water vapour, and daily patterns36. found that temperature, wind,
and cyclones can be utilised to forecast rain in growth of agriculture sector and the farmers can take their
decisions accordingly. Some researchers have also utilised atmospheric features such as temperature, relative
humidity, pressure, and wind speed to accurately predict rainfall using ML techniques such as ANN, RF,
and multiple linear regression models37,38. Therefore, the combination of ML, DL, and time series modeling
represents a promising frontier for advancing our understanding of climatic patterns and improving our ability
to predict rainfall.
This study focuses on exploring rainfall data in the North-Western Himalayas with several key objectives.
Firstly, we aim to gauge the efficacy of advanced ML algorithms in predicting rainfall patterns. Secondly, we
seek to evaluate the accuracy of DL models specifically tailored for this geographical region. Thirdly, we intend
to compare the performance of ML and DL methods against traditional time series techniques commonly
used in meteorological predictions. Additionally, our study aims to analyze how variations in altitude affect the
precision of these predictive models. Lastly, based on our findings, we aim to propose a holistic approach to
enhance the overall accuracy of rainfall prediction in the region. Through these objectives, we aim to contribute
valuable insights that could advance both scientific understanding and practical applications in meteorological
forecasting in mountainous terrains. This study takes a multidimensional approach to unravel the complexities
of precipitation trends in a region known for its intricate altitude-dependent climate variations. By integrating
these innovative methodologies, we aim to decode the intricate relationships between environmental factors and
rainfall, ultimately developing more accurate and localised predictive models.

Materials and methods

Case study region, meteorological data and pre-processing. The study area covers the North-Western Himalayan
region of India, centered at approximately 33° 5′ 24′′ N latitude and 74° 47′ 24′′ E longitude. Figure 1 illustrates
the location of the study area covered. The study map was generated by QGIS 3.30.0. The six locations of
meteorological stations were located on map across the altitudinal gradient. The primary data was gathered
from the head office of the Indian Meteorological Department (IMD). The dataset encompassed daily data
for maximum and minimum temperature and rainfall spanning 40 years (1980–2020). The initial data pre-
processing stage involved several tasks: data conversion, handling missing values, encoding categorical variables,
and partitioning the dataset into training and testing sets. Since the raw data exhibited both missing values and
extreme values or outliers, corrective actions were taken. Missing values were replaced by not available (N.A).
Subsequently, the dataset underwent encoding procedures before being ready for experimentation. Relevant
attributes crucial for rainfall prediction were identified and isolated. The dataset was then divided, with 80%
designated for training and 20% for testing, serving as input for the models.

Models
This study centres on rainfall prediction through a combined approach involving ML, DL, and time series
methodologies. The analysis encompassed three distinct ML algorithms: ANN, SVR, and RF; three DL algorithms:
RNN, LSTM, and GRU; and two time series algorithms: ARIMA and TBATS. These various algorithms utilised
input variables moderately to strongly associated with rainfall, drawn from multiple environmental factors.The
model parameters and information about models is given in supplementary Table 1S. The study determined and

Scientific Reports | (2024) 14:27876 | https://doi.org/10.1038/s41598-024-77687-x 3

www.nature.com/scientificreports/

Fig. 1. Study area map (map was created using QGIS 3.30.0 https://www.qgis.org/).

reported the most effective models and algorithms by assessing performance using the RMSE and MAE metrics,
other accuracy assement like Bias and R2 of various models across the altitudes is given in Table 1S.

Machine learning (ML) algorithms

Three ML algorithms were utilized to predict the rainfall based on the behaviour of the data set. The algorithms
are discussed below:

Artificial neural network (ANN)

ANN is an ML model inspired by the brain’s structure and function. The architecture of an ANN typically
consists of an input layer, one or more hidden layers, and an output layer. The input layer receives and passes
the input data to the hidden layers. The hidden layers use weights and biases to transform the input data into a
new representation that is more suitable for the task. The output layer produces the final predictions based on
the information processed by the hidden layers. The weights and biases are learned during training, allowing the
model to make increasingly accurate predictions. The number of hidden layers, the size of the hidden layers, and
the activation functions used determine the capacity of the model and its ability to learn complex relationships
between inputs and outputs. In Fig. 2, yt is considered as a function of its lag values yt−1, yt−2, . . . , yt−k and
respective weights.
q 
i=1 β ij yt−i ) + ϵ t (1)
k
yt = a0 + j=1 aj g(β 0j +

where aj (j = 0,1, 2, . . . , q) and β ij (j = 0,1, 2, . . . , q; i = 1,2, ., k) are the model’s connection weights; k is
the number of nodes in the input layer, and q is the number of nodes in the hidden layer.

Support vector regression (SVR)

SVR, developed by Vapnik et al.39, is a type of ML model used for regression and time series problems. The
SVR model architecture consists of the following components. First, the input features are transformed into a
high-dimensional space using a kernel function, such as a radial basis function (RBF) or a polynomial function.
The transformed features are used to train a model. The model finds the optimal boundary by maximising the

Scientific Reports | (2024) 14:27876 | https://doi.org/10.1038/s41598-024-77687-x 4

www.nature.com/scientificreports/

Fig. 2. Model architecture of ANN.

margin between the data points and the boundary. The model parameters are estimated using a cost function
penalising deviations from the target values. The model makes predictions for new input data by transforming
the features into high-dimensional space and applying the linear regression model. The kernel function and the
regularisation parameter (C) are hyperparameters that can be tuned to improve the model’s performance. The
main advantage of the SVR model is that it can handle complex non-linear relationships between inputs and
outputs using a suitable kernel function.

Random forest (RF)

RF is an ML algorithm based on decision trees. It builds multiple decision trees and combines their predictions
to improve the overall accuracy and stability of the model. Many decision trees are trained on randomly selected
subsets of the training data. Each decision tree is trained to make predictions for the target variable by recursively
splitting the data into subsets based on the values of the input features. The final prediction for a new input
sample is the average of the predictions made by all the decision trees. The random selection of features at each
split helps to decorrelate the trees and reduce overfitting. The bootstrapped training data samples used to train
each tree help reduce the prediction variance. The architecture of the RF model is given in Fig. 3.

Deep learning (DL) algorithms

To estimate daily rainfall through current environmental data, three specific algorithms viz., RNN, LSTM, and
GRU were selected for the empirical investigation. Consequently, these three DL algorithms were tested and
juxtaposed to determine the superior methods for anticipating daily rainfall quantities.

Recurrent neural network (RNN)

An RNN is a DL model that processes sequential data. It has a feedback loop that allows information to be
passed from one sequence step to the next. The core architecture of an RNN consists of a repeating module with
a hidden state, which stores information about the previous time step, and an input-to-hidden layer, which takes
the current input and last hidden state as inputs and outputs the current hidden state. The hidden state is then
used as input to a hidden-to-output layer, which generates the output for the present time step. The weights of
the input-to-hidden and hidden-to-output layers are shared across all time steps, allowing the model to learn
patterns in the sequence data. The internal structure of an RNN model can be described using the following
equations:
Input Layer: The input layer receives the input sequence denoted as x(t), with t representing the time step.
Hidden Layer: The hidden layer contains the recurrent neuron that computes the hidden state h(t) using the
following equation:
h (t) = f (Whx*x( t) + Whh · h(t − 1) + bh)(2)

Scientific Reports | (2024) 14:27876 | https://doi.org/10.1038/s41598-024-77687-x 5

www.nature.com/scientificreports/

Fig. 3. Architecture of RF.

where, f is the activation function (e.g., tanh), Whx is the weight matrix for the input-to-hidden layer
connection, Whh is the weight matrix for the hidden-to-hidden layer connection, and bh is the bias vector.

Output Layer: The output layer computes the output y(t) using the following equation:
y (t) = Why * h (t) + by (3)

where, Why is the weight matrix for the hidden-to-output layer connection, and by is the bias vector.

These equations describe the basic structure of a simple RNN model. More complex RNN models may have
additional layers or use different types of activation functions or recurrent neurons. This architecture can be
extended with multiple hidden layers to form a deep RNN or multiple hidden states to create an LSTM or GRU.

Long short-term memory (LSTM) and its variant

LSTM is an RNN type widely used for processing sequential data such as time series, speech signals, and text.
It was introduced by40 as a solution to the vanishing gradient problem that occurs in traditional RNNs. The
LSTM model is an RNN type that can remember previous inputs for a long time. The architecture of an LSTM
network consists of memory cells, input gates, forget gates and output gates. The memory cells are responsible
for storing information for a prolonged period, whereas the gates control the flow of information in and out of
the memory cells. The input gate controls the flow of information into the memory cell, allowing the network to
decide what information to retain. The forget gate decides what information should be discarded based on the
current input and previous memory state. The output gate determines what information should be outputted
based on the current input and previous memory state. The structure of an LSTM unit can be described in the
following Fig. 4 and equations:
Input gate: Controls the amount of new information allowed into the cell state. It is calculated using the
following equation:
it = σ {Wi ∗ (ht−1 ∗ xt)} + bi(4)

where, it is the input gate at time step t, Wi is the weight matrix, ht−1 is the concatenation of the previously
hidden state t − 1 and the current input xt, and bi is the bias term. The sigmoid activation function σ ensures
that the input gate is between 0 and 1.

Forget gate: Controls the amount of information that will be forgotten from the cell state. It is calculated using
the following equation:
ft = σ {Wf ∗ (ht−1 ∗ xt)} + bf (5)

where, ft is the forget gate at time step t, Wf is the weight matrix, and bf is the bias term.

Cell state: A continuous memory of the model that stores information from the past. It is updated using the
following equation:
∼
Ct = ft ∗ Ct−1 + it∗ Ct(6)

Scientific Reports | (2024) 14:27876 | https://doi.org/10.1038/s41598-024-77687-x 6

www.nature.com/scientificreports/

Fig. 4. Structure of an LSTM unit.

where,
∼ Ct is the cell state at time step t, ft is the forget gate, it is the input gate, Ct−1 is the previous cell state,
and Ct is a candidate cell state calculated as follows:
∼
Ct= tanh {Wc ∗ (ht−1 ∗ xt)} + bc(7)

Wc and bc are the weight and bias of the candidate’s input.

Output gate: Controls the amount of information output from the cell state to the hidden state. It is calculated
using the following equation:
ot = σ {Wo ∗ (ht−1 ∗ xt)} + bo(8)

where, ot is the output gate at time step t, Wo is the weight matrix, and

bo is the bias term.

Hidden state: A representation of the current state of the model. It is calculated using the following equation:
ht = ot ∗ tanh (Ct)(9)

where, ht is the hidden state at time step t, ot is the output gate, and Ct is the cell state.

The above equations form the core of an LSTM unit, and multiple units can be stacked together to create a multi-
layer LSTM network. A deep LSTM model is a variant of the LSTM model with multiple layers of LSTM units. A
bidirectional LSTM processes the input sequences in two ways: in the forward direction (from start to end) and
backward order (from end to start).

Gated recurrent unit (GRU)

GRU is a modification of the RNN model with two gates: the reset gate and the update gate. These gates help
regulate the flow of information and prevent the vanishing gradient problem that can occur in traditional RNNs.
The architecture of a GRU is like an LSTM network but with fewer parameters.

Time series algorithms

ARIMA and TBATS were chosen to forecast daily rainfall intensity using real-time environmental data. These
algorithms were tested and compared to determine the best approach for accurate daily rainfall predictions.

Scientific Reports | (2024) 14:27876 | https://doi.org/10.1038/s41598-024-77687-x 7

www.nature.com/scientificreports/

Autoregressive integrated moving average (ARIMA)

ARIMA is a popular time series forecasting algorithm to model and predict univariate data. For the ARIMA
model, the time series data must be stationary, meaning its mean, variance, and covariance are constant over
time. If the data is not stationary, a differencing step is applied to make it stationary. The ARIMA model is
selected based on the values of p, d, and q, where p is the number of AR terms, d is the number of differencing
steps applied, and q is the number of MA terms. The autocorrelation function (ACF) and partial autocorrelation
function (PACF) are used to identify the patterns in the time series data and to determine the number of
autoregressive (AR) and moving average (MA) terms to include in the ARIMA model. The model is fitted to
the time series data using maximum likelihood estimation or a variation thereof. The model is validated using
residual analysis and by comparing the predicted values with the actual values for a portion of the time series
data set aside for validation. The ARIMA model predicts the future values of the time series data.

Exponential smoothing state space model with Box–Cox transformation, ARMA errors,
trend and seasonality components (TBATS)
TBATS is a time series forecasting algorithm for univariate time series data. It is a hybrid algorithm that
combines exponential smoothing and ARIMA models, making it suitable for modeling complex time series
patterns like seasonality, trend, and irregularity. The algorithm models the time series as a combination of
multiple components, including a trend, a seasonal component, and an irregular component. The algorithm
estimates the model’s parameters using maximum likelihood estimation and predicts future values by combining
the different elements. TBATS has been found to perform well on various time series datasets, including those
with multiple seasonal patterns, non-stationary trends, and irregular fluctuations.

Statistical evaluation
Generally, it is necessary to assess the performance of a developed prediction model and compare it with
other models using specific statistical measures. However, it is crucial to employ multiple statistical indices
because different models may yield similar or nearly identical values for a particular index. This similarity
makes it challenging to definitively determine which model performs better than the others. Each statistical
index evaluates the model’s performance from a single perspective of how well its outputs match the desired
values. Therefore, evaluating models across multiple statistical indices is advisable to comprehensively assess
each model’s performance and conduct a robust comparative analysis, ultimately identifying the most suitable
modeling approach. Two performance metrics were commonly used to evaluate the effectiveness of predictive
models: the RMSE, and MAE. RMSE measures the average magnitude of the errors between predicted and
actual values. Lower RMSE values indicate better model accuracy, as they signify smaller discrepancies between
predicted and observed data points. Similarly, MAE quantifies the average absolute difference between predicted
and actual values, with lower values indicating more accurate predictions. The RMSE and MAE can be defined as

1 n  
2
RMSE = t=1 Yt − Yt
n
1  n  
 t
MAE = t=1 Y t − Y
n

where Yt is the actual value, Yt is the fitted value and n is the number of observations.

Result and discussion

The descriptive statistics of temperature across altitudinal gradients in the North-Western Himalayas reveal
significant variability, which is crucial for understanding local climate dynamics and its implications for rainfall
patterns. Mean maximum temperatures ranged from 11.65 to 20.21 ℃ across locations L1 to L6, reflecting
diverse thermal regimes influenced by altitude (Table 1). These findings are consistent with previous studies
emphasizing the role of elevation in temperature variation41. The wide range of maximum temperature values,
from − 5.70 to 37.60 ℃, reveals the broad climatic spectrum within the study area, indicative of both alpine
and lower altitude climates. Similarly, mean minimum temperatures ranged from 2.50 to 7.64 ℃ across the
same locations, highlighting cooler conditions at higher elevations compared to lower lying areas (Table 2).
This gradient is critical for understanding temperature-dependent processes such as evapotranspiration and
precipitation formation42. Standard error and skewness values for both maximum and minimum temperatures
further elucidate the distribution characteristics, indicating slight asymmetry in temperature data distributions
but with relatively low variability, suggesting robustness in measurements34,43,44. These temperature statistics
serve as a foundational framework for predicting rainfall patterns in the region. Studies have shown that
temperature gradients directly influence atmospheric stability, moisture content, and the onset of precipitation
events45–48. By integrating altitude-specific temperature data into rainfall models, researchers can enhance the
accuracy of predictive models tailored to mountainous regions49.
The analysis of rainfall patterns across different altitudinal locations (L1 to L6) in the North-Western Himalayas
provided valuable insights into the spatial variability and predictive challenges associated with precipitation
modeling in mountainous regions. Mean rainfall values ranging from 1.94 to 4.09 mm illustrate the significant
diversity in precipitation levels across relatively small geographical distances (Table 3). This variability is further
underscored by the wide range of minimum and maximum rainfall values observed at each location, reflecting
the complex interplay of topography, atmospheric dynamics, and local climatic conditions. The standard error
and skewness values offer additional perspectives on the distributional characteristics of rainfall data. Higher

Scientific Reports | (2024) 14:27876 | https://doi.org/10.1038/s41598-024-77687-x 8

www.nature.com/scientificreports/

Maximum temperature
Descriptive statistics L1 L2 L3 L4 L5 L6
Mean 18.62 16.60 20.02 19.31 11.65 20.21
Standard error 0.07 0.07 0.07 0.07 0.07 0.08
Median 19.96 17.90 21.10 20.70 12.22 21.60
Mode 28.00 25.00 29.00 27.50 21.60 30.50
Standard deviation 8.76 8.35 8.92 8.44 7.96 9.30
Kurtosis − 1.11 − 1.14 − 1.07 − 1.00 − 1.19 − 1.13
Skewness − 0.29 − 0.29 − 0.28 − 0.37 − 0.15 − 0.28
Range 41.00 40.20 43.60 40.80 38.70 43.30
Minimum − 5.70 − 8.00 − 6.60 − 5.10 − 9.60 − 5.70
Maximum 35.30 32.20 37.00 35.70 29.10 37.60

Table 1. Descriptive statistics of maximum temperature at different altitudinal gradients in the North-Western
Himalayas.

Minimum temperature
Descriptive statistics L1 L2 L3 L4 L5 L6
Mean 6.48 3.04 7.64 6.44 2.50 6.39
Standard error 0.06 0.06 0.06 0.06 0.06 0.06
Median 6.40 3.00 7.40 6.10 2.80 6.20
Mode 0.40 5.00 0.00 0.00 12.40 0.40
Standard deviation 6.95 7.09 7.51 7.12 7.20 7.19
Kurtosis − 0.98 − 0.57 − 1.11 − 0.94 − 1.08 − 1.05
Skewness 0.01 − 0.07 0.09 0.08 − 0.11 0.12
Range 38.60 64.60 37.00 36.40 37.90 39.10
Minimum − 15.70 − 18.60 − 11.80 − 13.60 − 19.80 − 15.70
Maximum 22.90 46.00 25.20 22.80 18.10 23.40

Table 2. Descriptive statistics of minimum temperature at different altitudinal gradients in the North-Western
Himalayas.

Rainfall
Descriptive statistics L1 L2 L3 L4 L5 L6
Mean 2.98 3.51 1.94 3.32 4.09 2.97
Standard Error 0.07 0.07 0.05 0.09 0.08 0.06
Median 0.00 0.00 0.00 0.00 0.00 0.00
Mode 0.00 0.00 0.00 0.00 0.00 0.00
Standard Deviation 8.26 8.70 6.30 10.37 10.13 7.65
Kurtosis 44.80 33.66 56.41 53.54 49.59 21.75
Skewness 5.46 4.77 6.18 6.02 5.61 4.14
Range 149.50 138.90 130.30 206.00 189.20 85.00
Minimum 0.00 -5.50 0.00 0.00 -0.20 0.00
Maximum 149.50 133.40 130.30 206.00 189.00 85.00

Table 3. Descriptive statistics of rainfall at different altitudinal gradients in the North-Western Himalayas.

skewness values, such as those observed for locations L2, L3, L4, and L5, indicate asymmetric distributions with a
tendency towards higher or lower rainfall extremes, potentially influenced by localized weather phenomena like
orographic lifting or convective processes44,50. Conversely, lower skewness values suggest a more symmetrical
distribution of rainfall, indicative of more moderate and consistent precipitation patterns51–54. The application
of ML models (RF, SVR, ANN, and KNN) to predict rainfall from meteorological variables reveals varying
model performances across different altitudinal gradients. RF and SVR models generally exhibited higher
RMSE values during both training and testing phases, suggesting challenges in capturing the complex nonlinear
relationships between predictors and rainfall outcomes in mountainous terrain. These findings align with
previous studies emphasizing the sensitivity of statistical models to spatial heterogeneity and the need for robust

Scientific Reports | (2024) 14:27876 | https://doi.org/10.1038/s41598-024-77687-x 9

www.nature.com/scientificreports/

validation strategies in mountainous regions55–57. In contrast, ANN and KNN models demonstrated relatively
lower RMSE values, indicating their potential for better capturing spatial variability and nonlinearity in rainfall
patterns across diverse altitudes. The superior performance of these models may stem from their ability to learn
complex patterns and relationships inherent in meteorological data, including altitude-dependent factors such
as orographic effects and microclimatic variations58,59. These findings highlight the importance of integrating
altitude-specific meteorological data and employing advanced modeling techniques to enhance the accuracy
of rainfall forecasts in mountainous regions. Future research directions could include refining model inputs by
incorporating additional environmental variables (e.g., terrain characteristics, vegetation cover) and exploring
ensemble modeling approaches to mitigate uncertainties associated with individual model performances60–64.
Furthermore, incorporating high-resolution satellite data and ground-based observations could further improve
the spatial representation of rainfall patterns and support more robust model validations in complex terrain
settings.
The application of ML, DL, and time series modeling techniques to predict rainfall from meteorological
variables across altitudinal gradients in the North-western Himalayas offers valuable insights into model
performance and predictive accuracy (Table 4). ML models such as RF, SVR, ANN, and KNN exhibited varying
levels of effectiveness as indicated by their train and test RMSE values across altitudinal levels (L1 to L6) (Fig. 5).
These differences underscored the sensitivity of these models to spatial variations in climatic conditions and
the complex interactions between meteorological predictors and rainfall patterns. DL models, including LSTM,
bi-directional LSTM, deep LSTM, GRU, and RNN, also demonstrated varying RMSE values across altitudinal
gradients. These models leverage sequential dependencies in data and have shown promise in capturing temporal
patterns and non-linear relationships in climatic variables, which are crucial for accurate rainfall prediction
in dynamic mountainous environments40,65. Moreover, time series modeling approaches such as ARIMA and
TBATS provided insights into the temporal variability of rainfall across altitudes. These models exhibited varying
train and test RMSE values, reflecting their ability to capture both short-term fluctuations and long-term trends
in rainfall data56,66–68. The observed differences in RMSE values across these modeling techniques highlight the
importance of selecting appropriate methodologies that account for the complex spatial and temporal dynamics
inherent in mountainous regions. The higher RMSE values in some models indicate challenges in accurately
capturing local-scale variations and extreme weather events, which are critical for effective water resource
management and disaster preparedness in mountain ecosystems69–71. Integrating diverse modeling approaches
and leveraging advanced statistical techniques enhance our understanding of rainfall variability in mountainous
areas.
The DL models, including LSTM, RNN, Bidirectional LSTM, deep LSTM, and GRU, exhibited varying degrees
of effectiveness across the study locations. LSTM consistently demonstrated robust performance, achieving lower
RMSE and MAE compared to other DL models in several instances30. For example, at Location 1, LSTM achieved
a training RMSE of 26.4051 and a testing RMSE of 38.2848, indicating its proficiency in capturing the complex
relationships between temperature variables and rainfall patterns. Bidirectional LSTM also showed competitive
performance, particularly noteworthy for its lower MAE scores across multiple locations. However, it displayed
marginally higher RMSE values compared to LSTM, suggesting potential variability in predictive accuracy
across different evaluation metrics. On the other hand, Deep LSTM and GRU, while generally performing
adequately, exhibited higher RMSE and MAE values compared to LSTM and Bidirectional LSTM (Fig. 6).
This observation suggests that these models might have struggled more with capturing the intricate nuances of
temperature-rainfall relationships specific to the north-western Himalayan region. In contrast, traditional ML
models such as ANN, KNN, SVR, RF, and the time series approach ARIMA consistently demonstrated higher
RMSE and MAE values across all locations72. Specifically, KNN, SVR, and RF exhibited notably higher error
metrics, highlighting their limitations in accurately capturing the non-linear dependencies inherent in rainfall
prediction tasks compared to DL models.
The DL models consistently outperformed both ML and time series models across all six locations examined
(Fig. 7). Specifically, LSTM and Bidirectional LSTM emerged as the top-performing models, achieving the
lowest RMSE and MAE scores for locations L1 through L6, respectively. Meanwhile, RNN, LSTM, Bidirectional
LSTM, and Deep LSTM were identified as superior models based on their performance in terms of test MAE
across the same locations. Overall, the DL algorithms demonstrated superior accuracy, with Bi-directional
LSTM showing the highest effectiveness, followed by LSTM, RNN, Deep LSTM, and GRU in descending order.
In contrast, the ML models performed relatively better than traditional time series methods, with ANN leading,
followed by KNN, SVR, and RF. Time series models, represented by TBATS and ARIMA, ranked lowest in
accuracy (Fig. 8). The preference for DL models in predicting rainfall patterns in the North-Western Himalayas
can be attributed to several factors. Firstly, the complex and non-linear nature of the data, influenced by altitude
variations, topography, and atmospheric stability, poses challenges for ML and time series models in accurately
capturing these relationships73. DL models, such as LSTM and Bidirectional LSTM, are specifically designed to
handle such complexities by automatically learning intricate data patterns27,72. This capability reduces the need
for manual feature engineering and enhances prediction accuracy by effectively managing noise inherent in
rainfall and temperature data. DL models benefit from the availability of large historical datasets in the region,
which are essential for training these models effectively. The extensive data enable DL models to generalize well
and remain robust against temporal variations in the data, thereby improving overall prediction accuracy. This
study makes a significant contribution to overcoming challenges in agricultural planning and climate adaptation
by applying advanced Machine Learning (ML) and Deep Learning (DL) techniques to rainfall prediction. As
traditional forecasting methods struggle with the increasing unpredictability of climate, this research enhances
forecast accuracy through sophisticated algorithms, allowing for more informed decisions by farmers and
agricultural planners. By integrating complex climatic data and assessing model effectiveness in the North-
Western Himalayas, the study provides crucial insights into improving predictive precision and managing varied

Scientific Reports | (2024) 14:27876 | https://doi.org/10.1038/s41598-024-77687-x 10

Scientific Reports |
(2024) 14:27876
www.nature.com/scientificreports/

L1 L2 L3
Models Train RMSE Test RMSE Train MAE Test MAE Train RMSE Test RMSE Train MAE Test MAE Train RMSE Test RMSE Train MAE Test MAE
LSTM 26.4051 38.2848 18.3032 22.272 28.6268 32.6531 18.6174 20.2858 20.9055 22.4477 14.0341 14.8548
RNN 26.4314 38.4852 17.9959 22.048 34.4668 36.1414 28.5762 28.7417 23.2679 24.2392 18.0615 18.1207
Bidirectional LSTM 28.5082 39.0882 20.9806 24.3876 28.4072 32.4625 18.7237 20.3433 20.738 22.2265 13.966 14.6357
Deep LSTM 29.2431 39.3847 22.2205 25.5191 30.2622 33.0527 22.1146 23.2161 22.7085 23.8022 17.0902 17.3248
GRU 29.5688 39.4416 22.7908 25.8663 31.0385 33.4672 23.8292 24.8423 22.0524 23.0137 14.7668 15.453
TBATS 30.4407 39.9763 23.7599 26.9549 33.7797 35.5372 26.5045 26.1142 23.6468 24.5349 17.885 17.2726
ANN 29.5231 40.6845 21.3692 23.5784 31.4014 37.3677 23.3202 27.4649 22.0548 24.0753 15.5666 15.8668
KNN 39.137 44.8777 26.0212 26.8164 41.1062 42.5591 29.3555 30.6486 29.0738 29.6286 19.4788 20.0675
SVR 37.0156 46.6131 25.3716 30.5223 40.2145 42.0613 29.3366 30.4928 28.7573 29.8899 19.6855 20.3037
RF 36.5499 48.5559 25.5494 31.5407 42.4912 43.1474 29.9277 29.323 30.6046 33.4445 19.9559 21.6242
ARIMA-X 33.3473 48.9682 24.8915 40.2202 33.7015 35.8058 27.2914 28.2475 23.2583 24.5984 17.7394 18.571

| https://doi.org/10.1038/s41598-024-77687-x
L4 L5 L6
Models Train RMSE Test RMSE Train MAE Test MAE Train RMSE Test RMSE Train MAE Test MAE Train RMSE Test RMSE Train MAE Test MAE
LSTM 37.6824 41.882 25.7388 23.2889 36.9767 32.2304 22.3422 21.5748 27.0726 27.1765 19.3609 20.128
RNN 34.9175 42.6456 21.6094 22.2904 36.6707 32.2461 22.3529 21.7823 27.7091 28.0215 19.5725 20.5426
Bidirectional LSTM 34.9437 42.6471 21.2045 21.8476 36.815 34.6105 25.4816 25.6676 26.5043 26.9585 18.4087 19.5685
Deep LSTM 36.629 44.0575 25.7735 26.5994 38.4883 35.9311 26.3701 26.5954 28.9267 29.2136 22.0771 22.8641
GRU 36.6263 44.1257 25.4547 26.1814 41.3881 35.9506 30.7579 27.3958 28.2329 28.6365 21.3745 22.0169
TBATS 39.4747 45.1084 30.271 29.8734 39.0508 37.0425 27.7653 28.1527 30.4776 30.9879 23.0755 24.0622
ANN 37.902 45.2663 28.3823 29.2586 39.0998 37.0643 27.8667 28.2929 26.633 30.0485 19.4238 20.9501
KNN 45.3611 51.0009 32.2122 38.5397 45.1007 40.61 31.9854 27.5973 37.01 37.7642 25.2801 27.1359
SVR 48.4088 52.6685 32.5023 33.006 47.412 44.517 32.5436 30.7088 36.3563 37.1531 25.3094 26.9675
RF 47.8864 53.3362 32.6218 32.8244 47.9481 45.5619 32.3674 31.2958 38.7174 37.8615 25.8642 25.9618
ARIMA-X 51.0779 58.2526 33.3209 35.5394 51.7974 53.9499 33.4704 34.7775 31.444 30.9085 24.5291 24.6846

Table 4. Performance metrics of predictive models at different altitudinal gradients in the North-Western Himalayas.

11
www.nature.com/scientificreports/

Fig. 5. Bar plot of RMSE across altitudinal gradient.

Fig. 6. Bar plot of MAE across altitudinal gradient.

environmental conditions. This advancement supports better agricultural practices, infrastructure development,
and overall food security in the face of climate variability.
The use of advanced Machine Learning (ML) and Deep Learning (DL) techniques for rainfall prediction,
as outlined in this study, represents a significant advancement in meteorological forecasting with important
implications for engineers and stakeholders, especially in semi-arid regions. Techniques such as Random Forest
(RF), Support Vector Regression (SVR), Long Short-Term Memory (LSTM), and Gated Recurrent Units (GRU)
offer notable improvements in rainfall prediction accuracy. This enhanced precision is crucial for effective
planning and management in areas dependent on seasonal rains. However, applying these techniques in practice
brings several considerations. A key issue is balancing the accuracy of the models with their complexity. While
these advanced methods can deliver highly accurate forecasts, they also introduce increased computational
demands and complexity. Engineers and stakeholders must assess whether the benefits of improved accuracy
justify the challenges associated with implementing and maintaining these sophisticated systems. The models

Scientific Reports | (2024) 14:27876 | https://doi.org/10.1038/s41598-024-77687-x 12

www.nature.com/scientificreports/

Fig. 7. Distribution of weather variables across altitudinal gradient.

Scientific Reports | (2024) 14:27876 | https://doi.org/10.1038/s41598-024-77687-x 13

www.nature.com/scientificreports/

Fig. 8. Taylor diagram of altitude wise comparing accuracy of models.

require significant computational power and specialized expertise, which might be a hurdle for regions with
limited technological resources. Proper integration of these advanced models into existing forecasting systems
is necessary to ensure that the accuracy improvements justify the operational demands. Another important
consideration is the dependency on extensive and high-quality datasets. These advanced models perform best
with large amounts of data, which may not always be available, especially in regions with sparse meteorological
information. This points to a need for better data collection infrastructure. Engineers should focus on enhancing

Scientific Reports | (2024) 14:27876 | https://doi.org/10.1038/s41598-024-77687-x 14

www.nature.com/scientificreports/

data acquisition systems and developing strategies to handle limited data to fully utilize these forecasting
techniques.
The adaptability of these methodologies to other meteorological variables like humidity, wind speed, and solar
radiation is also significant. While the principles behind these methods are versatile, each climatic variable has
unique challenges that may require specialized approaches. Although the methods used for rainfall prediction
can be extended to other variables, engineers need to tailor the models to each specific context to ensure
accuracy and effectiveness. Operational integration is another practical challenge. Incorporating these advanced
models into existing systems for agricultural planning, disaster management, and infrastructure development
requires careful consideration. Stakeholders must address how to integrate these models into practical decision-
making processes, which includes not only technical adaptation but also training and capacity building for users.
Ensuring that personnel are properly trained to use these advanced tools is essential for maximizing their benefits.
The benefits of improved rainfall forecasting are considerable. Enhanced prediction accuracy enables better
agricultural planning, such as more precise irrigation scheduling and effective water resource management. It
also supports infrastructure design and disaster preparedness by helping engineers create infrastructure that can
withstand extreme weather events. For policymakers, advanced forecasting techniques provide a foundation for
developing data-driven policies that promote sustainable resource management and enhance climate resilience.
In summary, while advanced techniques for rainfall prediction offer considerable advantages, they also require
careful management of model complexity, data needs, and operational integration. Addressing these challenges
effectively can lead to significant improvements in decision-making, infrastructure planning, and policy
development, ultimately supporting more resilient and sustainable practices in meteorology and related fields.
The model parameters were presented in Table 2S.

Conclusions
In conclusion, the comprehensive analysis of temperature and rainfall patterns across altitudinal gradients in
the North-Western Himalayas portrays the significant variability and complexity inherent in mountainous
climatic systems. The study revealed diverse thermal regimes influenced by altitude, with mean maximum
temperatures ranging from 11.65 to 20.21 ℃ and mean minimum temperatures from 2.50 to 7.64 ℃ across
different locations. Altitude is a critical factor shaping temperature variations in the region. The wide range
of temperature values reflects both alpine and lower altitude climates, crucial for understanding local climate
dynamics, evapotranspiration processes, and precipitation formation. Statistical analyses, including standard
error and skewness, further elucidate the distribution characteristics of temperature data, emphasizing
robustness in measurements despite slight asymmetry. This foundational framework of temperature statistics
not only enhances our understanding of regional climate but also serves as a crucial basis for predicting
rainfall patterns. Integrating altitude-specific temperature data into predictive models improves the accuracy
of rainfall forecasts by accounting for temperature gradients that influence atmospheric stability, moisture
content, and precipitation onset. Advanced modeling techniques such as ML, DL, and time series analysis
provided deeper insights into rainfall variability across diverse altitudinal gradients. DL models, particularly
LSTM and Bidirectional LSTM, demonstrated superior performance in capturing complex climatic relationships
compared to traditional ML and time series methods. Their ability to handle non-linear data dynamics and
leverage extensive historical datasets underscores their effectiveness in predicting rainfall patterns in dynamic
mountainous environments. Moving forward, continued research efforts should focus on refining model inputs
by incorporating additional environmental variables and exploring ensemble modeling approaches to further
enhance prediction accuracy. High-resolution satellite data and ground-based observations will play pivotal
roles in improving spatial representation and validating models in complex terrain settings. By advancing our
understanding and predictive capabilities, we can better manage water resources and mitigate risks associated
with climate variability in mountain ecosystems. In real-world applications for rainfall monitoring and warning
systems, standalone methods such as statistical, physical, and data-driven models offer distinct advantages but
also face notable limitations. Statistical models, like ARIMA, are straightforward and computationally efficient,
making them practical for immediate use. However, their tendency to assume linear relationships and their
limited flexibility can lead to suboptimal performance when faced with sudden climatic changes or complex
weather patterns. Physical models, which simulate atmospheric and hydrological processes, can provide in-
depth forecasts by considering intricate variable interactions. Despite their accuracy in well-defined conditions,
these models are often hindered by high computational demands and complexity, making them less feasible
for real-time applications in regions with limited resources. Additionally, physical models may also encounter
local optima if not precisely calibrated for specific regional conditions. Data-driven models, such as machine
learning and deep learning techniques, are adept at identifying complex, non-linear patterns and can be adapted
to various datasets. While they have the potential for high precision, their effectiveness depends on access to
extensive and high-quality data and significant computational power. These models are also prone to overfitting,
where they excel with historical data but may underperform with new or different data, and their intricate
nature often results in lower interpretability. Overall, while advanced data-driven methods offer substantial
improvements in rainfall forecasting, their practical implementation must carefully address data quality,
computational requirements, and ongoing maintenance to ensure reliable and actionable predictions.

Data availability
The datasets used and/or analyzed during the current study are available from the corresponding author on
reasonable request.

Received: 2 August 2024; Accepted: 24 October 2024

Scientific Reports | (2024) 14:27876 | https://doi.org/10.1038/s41598-024-77687-x 15

www.nature.com/scientificreports/

References
1. Skendžić, S., Zovko, M., Živković, I. P., Lešić, V. & Lemić, D. The impact of climate change on agricultural insect pests. Insects. 12,
440 (2021).
2. Zuma-Netshiukhwi, G., Stigter, K. & Walker, S. Use of traditional weather/climate knowledge by farmers in the South-Western
Free State of South Africa: agrometeorological learning by scientists. Atmosphere. 4, 383–410 (2013).
3. Jones, J. W., Hansen, J. W., Royce, F. S. & Messina, C. D. Potential benefits of climate forecasting to agriculture. Agric. Ecosyst.
Environ. 82, 169–184 (2000).
4. Chen, C. et al. Forecast of rainfall distribution based on fixed sliding window long short-term memory. Eng. Appl. Comput. Fluid
Mech. 16, 248–261 (2022).
5. Abobakr Yahya, A. S. et al. Water quality prediction model based support vector machine model for ungauged river catchment
under dual scenarios. Water. 11, 1231 (2019).
6. Grote, U. Can we improve global food security? A socio-economic and political perspective. Food Secur. 6, 187–200 (2014).
7. Trenberth, K. E. & Asrar, G. R. Challenges and opportunities in water cycle research: WCRP contributions. Earth’s Hydrol. Cycle
46, 515–532 (2014).
8. Franzke, C. L., O’Kane, T. J., Berner, J., Williams, P. D. & Lucarini, V. Stochastic climate theory and modeling. Wiley Interdiscip. Rev.
Clim. Change. 6, 63–78 (2015).
9. Wang, H. et al. Association of meteorological factors with infectious diarrhea incidence in Guangzhou, southern China: a time-
series study (2006–2017). Sci. Total Environ. 672, 7–15 (2019).
10. Garbrecht, J. D., Nearing, M. A., Zhang, J. X. & Steiner, J. L. Uncertainty of climate change impacts on soil erosion from cropland
in central Oklahoma. Appl. Eng. Agric. 32, 823–836 (2016).
11. Resnicow, K. & Vaughan, R. A chaotic view of behavior change: a quantum leap for health promotion. Int. J. Behav. Nutr. Phys.
Activity. 3, 1–7 (2006).
12. Praveen, B. et al. Analyzing trend and forecasting of rainfall changes in India using non-parametrical and machine learning
approaches. Sci. Rep. 10, 10342 (2020).
13. Khambra, G. & Shukla, P. Novel machine learning applications on fly ash based concrete: an overview. Mater. Today: Proc. 80,
3411–3417 (2023).
14. Alkesaiberi, A., Harrou, F. & Sun, Y. Efficient wind power prediction using machine learning methods: a comparative study.
Energies. 15, 2327 (2022).
15. Patil, S. S. & Vidyavathi, B. A machine learning approach to weather prediction in wireless sensor networks. Int. J. Adv. Comput.
Sci. Appl. 13 (2022).
16. Namitha, K., Jayapriya, A. & Kumar, G. S. In Proceedings of the Third International Symposium on Women in Computing and
Informatics. 492–495.
17. Wang, Z., Wang, Y., Zeng, R., Srinivasan, R. S. & Ahrentzen, S. Random forest based hourly building energy prediction. Energy
Build. 171, 11–25 (2018).
18. Levis, A. & Papageorgiou, L. Customer demand forecasting via support vector regression analysis. Chem. Eng. Res. Des. 83, 1009–
1018 (2005).
19. Calderaro, J., Seraphin, T. P., Luedde, T. & Simon, T. G. Artificial intelligence for the prevention and clinical management of
hepatocellular carcinoma. J. Hepatol. 76, 1348–1361 (2022).
20. Indrakumari, R., Poongodi, T. & Singh, K. Introduction to deep learning. Adv. Deep Learn. Eng. Sci. Pract. Approach. 1, 1–22
(2021).
21. Vieira, S., Pinaya, W. H. & Mechelli, A. Using deep learning to investigate the neuroimaging correlates of psychiatric and
neurological disorders: methods and applications. Neurosci. Biobehav. Rev.. 74, 58–75 (2017).
22. Janiesch, C., Zschech, P. & Heinrich, K. Machine learning and deep learning. Electron. Markets. 31, 685–695 (2021).
23. Ozcanli, A. K., Yaprakdal, F. & Baysal, M. Deep learning methods and applications for electrical power systems: a comprehensive
review. Int. J. Energy Res. 44, 7136–7157 (2020).
24. Kang, J. et al. Prediction of precipitation based on recurrent neural networks in Jingdezhen, Jiangxi Province, China. Atmosphere.
11, 246 (2020).
25. Mayer, M. J. Benefits of physical and machine learning hybridization for photovoltaic power forecasting. Renew. Sustain. Energy
Rev. 168, 112772 (2022).
26. Almási, A. D., Woźniak, S., Cristea, V., Leblebici, Y. & Engbersen, T. Review of advances in neural networks: neural design
technology stack. Neurocomputing. 174, 31–41 (2016).
27. Saha, S., Baral, S. & Haque, A. DEK-Forecaster: A novel deep learning model integrated with EMD-KNN for traffic prediction.
arXiv preprint arXiv:2306.03412 (2023).
28. Ban, W., Shen, L., Chen, J. & Yang, B. Short-term prediction of wave height based on a deep learning autoregressive integrated
moving average model. Earth Sci. Inf. 16, 2251–2259 (2023).
29. Wani, O. A. et al. Climate plays a dominant role over land management in governing soil carbon dynamics in North Western
Himalayas. J. Environ. Manage. 338, 117740 (2023).
30. Godahewa, R., Bergmeir, C., Webb, G. I. & Montero-Manso, P. An accurate and fully-automated ensemble model for weekly time
series forecasting. Int. J. Forecast. 39, 641–658 (2023).
31. Thirumalai, C., Harsha, K. S., Deepak, M. L. & Krishna, K. C. In 2017 International Conference on Trends in Electronics and
Informatics (ICEI). 1114–1117 (IEEE).
32. Prabakaran, S., Kumar, P. N. & Tarun, P. S. M. Rainfall prediction using modified linear regression. ARPN J. Eng. Appl. Sci. 12,
3715–3718 (2017).
33. Tharun, V., Prakash, R. & Devi, S. R. In 2018 Second International Conference on Inventive Communication and Computational
Technologies (ICICCT). 1507–1512 (IEEE).
34. Kumar, S. S., Wani, O. A., Krishna, J. R. & Hussain, N. Impact of climate change on soil health. Int. J. Environ. Sci. 7, 70–90 (2022).
35. Abhishek, K., Kumar, A., Ranjan, R. & Kumar, S. In 2012 IEEE Control and System Graduate Research Colloquium. 82–87 (IEEE).
36. Chaudhari, M. & Choudhari, D. Study of various rainfall estimation & prediction techniques using data mining. Am. J. Eng. Res. 6,
137–139 (2017).
37. Vijayan, R., Mareeswari, V., Mohankumar, P., Gunasekaran, G. & Srikar, K. Estimating rainfall prediction using machine learning
techniques on a dataset. Int. J. Sci. Technol. Res. 9, 440–445 (2020).
38. Gnanasankaran, N. & Ramaraj, E. A multiple linear regression model to predict rainfall using Indian meteorological data. Int. J.
Adv. Sci. Technol. 29, 746–758 (2020).
39. Vapnik, V. & Chervonenkis, A. The necessary and sufficient conditions for consistency in the empirical risk minimization method.
Pattern Recognit. Image Anal. 1, 283–305 (1991).
40. Hochreiter, S. & Schmidhuber, J. Long short-term memory. Neural Comput. 9, 1735–1780 (1997).
41. Palazzi, E., Filippi, L. & von Hardenberg, J. Insights into elevation-dependent warming in the Tibetan Plateau-Himalayas from
CMIP5 model simulations. Clim. Dyn. 48, 3991–4008 (2017).

Scientific Reports | (2024) 14:27876 | https://doi.org/10.1038/s41598-024-77687-x 16

www.nature.com/scientificreports/

42. Patel, J. B. Analysis of the microbial diversity associated with the Lesotho highlands through culture-independent approaches
(2020).
43. Dunn, R. J., Willett, K. M. & Parker, D. E. Changes in statistical distributions of sub-daily surface temperatures and wind speed.
Earth Sys. Dyn. 10, 765–788 (2019).
44. Nazir, S. F., Singh, L., Shah, B. A. & Ali, O. R. I. Rice-wheat cropping system under changing climate scenario: a review. Int. J. Chem.
Stud. 8, 1907–1914 (2020).
45. Gimeno, L. et al. Oceanic and terrestrial sources of continental precipitation. Rev. Geophys. 50 (2012).
46. Moharana, L., Sahoo, A. & Ghose, D. In IOP Conference Series: Earth and Environmental Science. 012054 (IOP Publishing).
47. Sahoo, A., Behera, S. & Sharma, N. in AIP Conference Proceedings. (AIP Publishing).
48. Sahoo, A. & Ghose, D. K. in Smart Intelligent Computing and Applications, Volume 1: Proceedings of Fifth International Conference
on Smart Computing and Informatics (SCI 307–317 (Springer, 2021).
49. Körner, C. et al. Creative use of mountain biodiversity databases: the Kazbegi research agenda of GMBA-DIVERSITAS. Mt. Res.
Dev. 27, 276–281 (2007).
50. Hornberger, G. M., Wiberg, P. L., Raffensperger, J. P. & D’Odorico, P. Elements of Physical Hydrology (JHU, 2014).
51. Rashid, M. M., Beecham, S. & Chowdhury, R. K. Statistical characteristics of rainfall in the Onkaparinga catchment in South
Australia. J. Water Clim. Change. 6, 352–373 (2015).
52. Sahoo, A. & Ghose, D. K. Imputation of missing precipitation data using KNN, SOM, RF, and FNN. Soft. Comput. 26, 5919–5936
(2022).
53. Sahoo, B. B., Jha, R., Singh, A. & Kumar, D. Long short-term memory (LSTM) recurrent neural network for low-flow hydrological
time series forecasting. Acta Geophys. 67, 1471–1481 (2019).
54. Sahoo, B. B., Panigrahi, B., Nanda, T., Tiwari, M. K. & Sankalp, S. Multi-step ahead urban water demand forecasting using deep
learning models. SN Comput. Sci. 4, 752 (2023).
55. Dal Molin, M., Schirmer, M., Zappa, M. & Fenicia, F. Understanding dominant controls on streamflow spatial variability to set up
a semi-distributed hydrological model: the case study of the Thur catchment. Hydrol. Earth Syst. Sci. 24, 1319–1345 (2020).
56. Singh, G., Batra, N., Salaria, A., Wani, O. A. & Singh, J. Groundwater quality assessment in Kapurthala district of central plain zone
of Punjab using hydrochemical characteristics. J. Soil Water Conserv. 20, 43–51 (2021).
57. Babu, S. et al. Biochar implications in cleaner agricultural production and environmental sustainability. Environ. Science: Adv. 2,
1042–1059 (2023).
58. Dura, V., Evin, G., Favre, A. C. & Penot, D. Spatial variability in the seasonal precipitation lapse rates in complex topographical
regions–application in France. Hydrol. Earth Syst. Sci. 28, 2579–2601 (2024).
59. Altaf, S. et al. Management of green mold disease in white button mushroom (Agaricus Bisporus) and its yield improvement. J.
Fungi. 8, 554 (2022).
60. Fisher, R. A. & Koven, C. D. Perspectives on the future of land surface models and the challenges of representing complex terrestrial
systems. J. Adv. Model. Earth Syst. 12, eMS001453 (2018).
61. Maxwell, A. E. & Shobe, C. M. Land-surface parameters for spatial predictive mapping and modeling. Earth Sci. Rev. 226, 103944
(2022).
62. Sahoo, B. B., Sankalp, S. & Kisi, O. A novel smoothing-based deep learning time-series approach for daily suspended sediment load
prediction. Water Resour. Manage. 37, 4271–4292 (2023).
63. Satapathy, D. P., Swain, H., Sahoo, A., Samantaray, S. & Satapathy, S. C. In Intelligent System Design: Proceedings of INDIA 2022
355–364 (Springer, 2022).
64. Swagatika, S., Paul, J. C., Sahoo, B. B., Gupta, S. K. & Singh, P. Improving the forecasting accuracy of monthly runoff time series of
the Brahmani River in India using a hybrid deep learning model. J. Water Clim. Change. 15, 139–156 (2024).
65. Greff, K., Srivastava, R. K., Koutník, J., Steunebrink, B. R. & Schmidhuber, J. LSTM: a search space odyssey. IEEE Trans. Neural
Netw. Learn. Syst. 28, 2222–2232 (2016).
66. Momani, P. & Naill, P. Time series analysis model for rainfall data in Jordan: Case study for using time series analysis. Am. J.
Environ Sci. 5, 599 (2009).
67. Bouznad, I. E. et al. Trend analysis and spatiotemporal prediction of precipitation, temperature, and evapotranspiration values
using the ARIMA models: case of the Algerian highlands. Arab. J. Geosci. 13, 1281 (2020).
68. Hasan, N. A., Dongkai, Y. & Al-Shibli, F. SPI and SPEI drought assessment and prediction using TBATS and ARIMA models,
Jordan. Water. 15, 3598 (2023).
69. Yucel, I. & Onen, A. Evaluating a mesoscale atmosphere model and a satellite-based algorithm in estimating extreme rainfall
events in northwestern Turkey. Nat. Hazards Earth Syst. Sci. 14, 611–624 (2014).
70. Kara, F., Yucel, I. & Akyurek, Z. Climate change impacts on extreme precipitation of water supply area in Istanbul: use of ensemble
climate modelling and geo-statistical downscaling. Hydrol. Sci. J. 61, 2481–2495 (2016).
71. Jamei, M. et al. Quantitative improvement of streamflow forecasting accuracy in the Atlantic zones of Canada based on hydro-
meteorological signals: a multi-level advanced intelligent expert framework. Ecol. Inf. 80, 102455 (2024).
72. Saha, S. et al. Measuring landslide vulnerability status of Chukha, Bhutan using deep learning algorithms. Sci. Rep. 11, 16374
(2021).
73. Huffman, G. J. & Bolvin, D. T. TRMM and other data precipitation data set documentation. NASA Greenbelt USA. 28, 1 (2013).

Acknowledgements
The authors thank the Researchers Supporting Project number (RSPD2024R730), King Saud University, Riyadh,
Saudi Arabia. Authors appreciated the support of the Division of Agronomy SKUAST Kashmir and NICRA,
project SKUAST Kashmir and IMD India for providing data.

Author contributions
Conceptualization, supervision, methodology, formal analysis, writing—original draft preparation, writing—
review and editing, O.A.W., S.S.M., M.Y., S.S.K., A.S.G.; data curation, project administration, investigation,
writing—review and editing, F.D., N.A.-A., S.E.-H., M.A.M. All authors have read and agreed to the published
version of the manuscript.

Funding
Open access funding provided by Lulea University of Technology. Researchers Supporting Project number
(RSPD2024R730), King Saud University, Riyadh, Saudi Arabia.

Scientific Reports | (2024) 14:27876 | https://doi.org/10.1038/s41598-024-77687-x 17

www.nature.com/scientificreports/

Declarations

Competing interests
The authors declare no competing interests.

Additional information
Supplementary Information The online version contains supplementary material available at ​h​t​t​p​s​:​/​/​d​o​i​.o
​ ​r​g​/​1​
0​.​1​0​38​ ​/​s​4​1​5​9​8​-​02​ ​4​-​7​7​6​8​7​-​x​​​​.​ ​​
Correspondence and requests for materials should be addressed to S.S.M., N.A.-A. or M.A.M.
Reprints and permissions information is available at www.nature.com/reprints.
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which
permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give
appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and
indicate if changes were made. The images or other third party material in this article are included in the article’s
Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included
in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or
exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy
of this licence, visit http://creativecommons.org/licenses/by/4.0/.

© The Author(s) 2024

Scientific Reports | (2024) 14:27876 | https://doi.org/10.1038/s41598-024-77687-x 18