PROJECT REPORT

ON
“RAINFALL DOWNSCALING”
STATISTICAL AND TEMPORAL
SUBMITTED IN FULFILMENT OF REQUIREDMENT FOR AWARD OF
BACHELOR OF ENGINEERING IN
DEPARTMENT OF CIVIL ENGINEERING
SUBMITTED BY
ANSHIKA GUPTA SAHIL BALHARA
YASH MEHTA
12111105 12111131
12111156
SUBMITTED TO
DR. KK SINGH
CIVIL DEPARTMENT

DEPARTMENT OF CIVIL ENGINEERING

NATIONAL INSTITUTE OF TECHNOLOGY
KURUKSHETRA
 DECLARATION
WE hereby declare that the seminar report “RAINFALL
DOWNSCALING” submitted for partial fulfilment of
requirements for the Award of Degree of “Bachelor of
Technology” in Department of Civil Engineering and
submitted to the Department of civil engineering,
National Institute of Technology, Kurukshetra is a record
of my own investigations carried under the Guidance of
Prof. KK SINGH Sir, Department of Civil Engineering. WE
have not submitted the matter presented in this Report
anywhere for the Award of any other Degree.
 ACKNOWLEDGEMENT

With great pride I would like to express my gratitude to

National Institute of Technology , Kurukshetra. The temple
of learning for providing us the required platform to
fulfilment of the project.

I express wholehearted gratitude to Dr. K.K Singh, Professor

of Civil Engineering Department. We wish to acknowledge
the support for making our task easy by providing us with
all valuable help and encouragement .

Finally, We would like to thank our Project mentor for his

kind co-operation and motivation to proceed in our project
work.

INTRODUCTION
Rainfall is a critical climatic variable that significantly
influences water resources, agriculture, ecosystem
health, and urban planning. Understanding and
predicting rainfall patterns are essential for managing
these resources, especially in the context of climate
change, which is altering precipitation trends
worldwide. However, climate models typically operate
at large spatial scales (regional or global), and their
coarse resolution can limit their usefulness for
applications at finer spatial scales, such as local or
watershed-level planning. This challenge has led to the
development of rainfall downscaling techniques,
which aim to derive high-resolution precipitation data
from coarser climate model outputs.
Rainfall downscaling is broadly classified into
dynamical and statistical methods. Dynamical
downscaling involves using high-resolution regional
climate models (RCMs) to simulate local climate
features based on global model outputs. While RCMs
are highly detailed, they are computationally expensive
and require substantial data and resources. In contrast,
statistical downscaling is a more computationally
efficient method that establishes a relationship
between large-scale atmospheric variables (e.g.,
temperature, pressure, wind patterns) and local
precipitation. These relationships are typically derived
using historical climate data and can be applied to
future climate scenarios generated by global circulation
models (GCMs).
Statistical downscaling techniques are particularly
valuable in areas where high-resolution climate models
are either unavailable or impractical to use. These
methods can be broadly categorized into two main
approaches: linear and non-linear models. Linear
models, such as multiple linear regression and
principal component analysis, assume that the
relationship between large-scale predictors and local
rainfall is linear. On the other hand, non-linear
methods, like neural networks or support vector
machines, can capture more complex, non-linear
relationships in the data, offering greater flexibility in
modeling rainfall behavior under varying climatic
conditions.
Another key aspect of rainfall downscaling is the
temporal scale of the data. Precipitation patterns
exhibit varying temporal scales, from short-term storm
events to seasonal and decadal trends. Downscaling
methods must account for these temporal dynamics
to produce reliable forecasts. Temporal downscaling
involves disaggregating coarse-scale rainfall data into
finer time steps, such as daily or hourly values, while
maintaining the statistical properties of the original
dataset. This aspect is crucial for applications like flood
modeling, drought prediction, and water resource
management.
Overall, rainfall downscaling, particularly through
statistical and temporal approaches, is an essential tool
for enhancing the resolution of climate projections. By
improving the accuracy and granularity of precipitation
data, these techniques provide valuable insights for
local climate adaptation and planning.

OBJECTIVE
THE STUDY AIMS TO PROVIDE HIGH RESOLUTION CLIMATE CHANGE
PROJECTION FOR KURUKSHETRA(HARYANA) IN TERMS OF
PRECIPITATION. THE GOAL IS TO GENERATE DETAILED REGIONAL
CLIMATE DATA(PRECIPITATION) TO BETTER UNDERSTAND THE
CLIMATE CHANGE IN TERMS OF PRECIPITATION USING GOOGLE
EARTH ENGINE CODE EDITOR , SDSM MODEL, PYTHON LANGUAGE
ETC.
USING STATISTICAL DOWNSCALING MODEL (SDSM) WE ANALYZE THE
PREVIOUS RAINFALL DATA TO PREDICT FUTURE RAINFALL CONDITION
FROM THAT IS FROM DIFFERENT SPECTRUM OF YEAR DAILY.
GOOGLE EARTH ENGINE USED CERTAIN MODELS TO CORRELATE THE
INDEPENDENT VARIABLE (SUCH AS VEGETATION, LANDCOVER,
DIGITAL ELEVATION ETC.) TO DEPENDENT VARIBLE THAT IS
PRECIPITATION TO FORM A MODEL THAT PREDICTS THE DATA WITH
1KM RESOLUTION FOR ONE YEAR MONTHLY BASIS
PYTHON AND TIME SERIES ANALYSIS ARE DONE TO PREDICT
RAINFALL DATA FOR 45 DAYS.
 REVIEW OF LITERATURE
Rainfall downscaling is an important methodology for
bridging the gap between large-scale climate models
(such as global circulation models, GCMs) and the local-
scale, high-resolution data needed for various
hydrological, agricultural, and urban planning
applications. Downscaling techniques allow for a more
detailed representation of rainfall patterns at a finer
spatial and temporal resolution, which is particularly
crucial for managing water resources, designing flood
control systems, and assessing climate change
impacts. There are two primary approaches to rainfall
downscaling: statistical downscaling and dynamical
downscaling, each with its own strengths, limitations,
and applications. Hence research paper of certain
authors were analysed to take a grip on concept.
AUTHOR TITLE SUMMARY
Kumar et.al Climate change In this research paper they
2014 projections for have downscaled the data
Tamil Nadu, using IMD and GCM as
India: deriving predicators and taking
high-resolution 1970-
climate data by a 2000 as baseline and to get
downscaling resolution of 25km*25km .
approach using PRECIS is a third generation
PRECIS Hadley RCM and its had
performed well in
estimating
the rainfall and
temperature
variation in Bangladesh and
India. The future
assessments
were performed by ignoring
model biases using:
(Model future + bias) -
(Model baseline + bias)=
change
HadCM3Q0 projected the
 closest estimation to IMD
data.
For next 100 years (2000-
2100) the maximum
temperature and minimum
temperature are expected
to
rise by 0.8°-3.5° and 3.1°
respectively. Rainfall
projections show no
significant
change by the end of the
century, but a slight
increase
during the northeast
monsoon
season.
Software used:
The selection was based on
evaluating the root mean
square errors (RMSE) of the
models in simulating
historical
rainfall data.
PRECIS, HadCM3Q0,

HadCM3Q1, HadCM3Q5,
HadCM3Q7, HadCM3Q11,
and HadCM3Q13.
Kumar et.al Downscaling In this study we have
2011c algorithms for compared deep learning
CMIP6 GCM based algorithms namely
daily rainfall SRCNN, stacked SRCNN and
over India DeepSD. Downscaling
meteorological data
involves
dynamical and statistical
approaches. Dynamical
downscaling uses high-
resolution regional models,
while statistical downscaling
uses global model outputs
 or
observations as inputs to
statistical models. Artificial
Intelligence and Machine
Learning are used today for
statistical downscaling and
Deep learning is a result of
it.
The SRCNN model is trained
on low-resolution rainfall
data
and can reconstruct high-
resolution images. Stacked
SRCNNs can boost the
resolution to a high degree
by
improving the resolution by
one factor at each step. The
DeepSD method uses
topography as an additional
input. The models were
developed using daily IMD
gridded data and tested on
five years of monsoon
season
data. DeepSD outperforms
other downscaling methods
based on pattern
correlation
and mean square error.
AI/ML-based postprocessing

and downscaling are

proposed
as efficient methods for
rainfall
data over the Indian region.
Software used:
SRNCC, stacked SRCNN,
DeepSD.
Raj et.al Downscaling This study evaluates the
2024 algorithms for effectiveness of various
CMIP6 GCM large-scale predictors in
daily rainfall reproducing local-scale
over India rainfall in India using
artificial neural networks,
change-factors, K-nearest
neighbour , and multiple
linear regression and also
contributes to selecting
suitable predictors and
comparing downscaling
algorithms for daily rainfall
in India. NCEP/NCAR
reanalysis data was used
to predict rainfall in India
from 1979 to 2014.
Predictors were selected
based on their correlation
with precipitation. The
NCEP/NCAR reanalysis
data was bias-corrected
and downscaled using
linear scaling and bilinear
interpolation. The
Change-Factors (CF)
method showed the best
performance in
reproducing local-scale
rainfall statistics with high
R² values (0.92–0.99),
low RMSE (1.37–28.88
mm), and NSE values
(–16.55–0.99).
- The CF approach
closely matched the
probability distribution

patterns observed in IMD

data.
Software used:
Artificial Neural Networks
Change-Factors
K-Nearest Neighbor
 Multiple Linear
Regression
Singh et.al Statistical In this research paper they
2015 downscaling and have downscaled of satluj river
projection of basin .
future It shows us the temperature
temperature and change ,flow change if the
precipitation river and other factors.
change in middle The study aims to compare
catchment of three downscaling techniques
Sutlej River to simulate maximum and
Basin, India minimum temperatures in the
Sutlej River basin, Central
India.
The simulated data of the
Canadian Centre for Climate
Modelling and Analysis
CGCM3 T47 version were
used for future climate
projects
Percentage reduction in partial
correlation
(PRP) for each predictor is
derived from the
Equation given below:
PRP = (P r – R1)/R1
Software used:
Software used in this is SDSM
model
A comparison between
downscaled results
Of CGCM3 model and
HadCM3 model is also made
Under a common emission
scenario,

Kumar et al2024 Evaluating

future urban
temperature
over smart
cities of the
Gangetic
 plains using
statistically
downscaled
CMIP6

This study tells us about the

future temperature and
extreme
temperature indices over
smart
cities in the Gangetic Plains
using downscaled CMIP6
model
outputs.
Results show that these cities

projections are likely to

experience

warmer
and more extreme
temperatures
in the upcoming decades.
The “Middle of the road”
scenario
(SSP245) suggests moderate
growth and slow sustainability
progress, resulting in 4.5
W/m2
radiative forcing.
Software used:
GCM,SDMS,IMD
Kumar et Evaluating The paper presents a spatial-
al2024 future urban temporal downscaling method
temperature to enhance Intensity-Duration-
over smart Frequency (IDF) relationships
cities of the by integrating large-scale
Gangetic climate data from General
plains using Circulation Models (GCMs)
statistically with local extreme
downscaled precipitation records. The
CMIP6 method involves spatial
 downscaling through the SD
Rain model, linking GCM
simulations to local daily
precipitation, and temporal
downscaling using the General
Extreme Value (GEV)
distribution to connect daily
and sub-daily precipitation.
The approach was tested with
data from Quebec, Canada,
and South Korea (1961-2000),
applied to both historical data
and future climate scenarios
using the Canadian CGCM3
and UK HadCM3 models.
Findings showed improved
accuracy after bias correction
and demonstrated that sub-
daily rainfall can be predicted
from daily GCM data, though
GCMs may provide differing
projections of future rainfall
trends. The downscaling
method proved effective for
constructing IDF curves,
essential for climate impact

assessments.
Software used
SD RAIN, GCM (CANADIAN
CGCM3 AND UK HADCM3)
 Section I
Rainfall downscaling through SDSM software
Statistical Downscaling Model (SDSM) Tutorial
 It will explain how to analyze previous rainfall data
to predict future rainfall conditions.
 The probable rainfall conditions after 30, 60, and
90 years will be discussed in comparison to the
present condition
Data type
 The SDSM software requires data to be in a

specific format for analysis.

 The primary data, such as temperature or

rainfall data, should be in a ".dht" file format,

and a predicting file named "box033x4toy" is
also needed
SDSM software
 Open SDSM software version 4.2, though

versions 5 or above, or below, can also be

used
 Change the settings by selecting the day of

the year, for example, 365 days

 Select the work period for analysis, such as

from the year 2000 to 2015

 Ensure the number of years in the work period

is odd
 In the miscellaneous settings, untick the

option to allow negative values

 Set the threshold for rainfall events to one

 For missing data, use the identifier -999

Missing data
 In the data, some days may have missing

data.
 Missing data should be changed to -999.
 If the data is zero, it is acceptable, but missing

data must be marked as -999.

 The software identifies missing data using the

-999 marker.
 After marking missing data, proceed to

advanced settings.
 In advanced settings, set the model

transformation to the fourth root.

 Other settings should remain as usual.

 For rainfall temperature analysis, use the

ordinary least squares algorithm.

 Save the settings.

 Select the default file directory where all data

is stored.
 In this case, the directory is on the C drive,

desktop.
 Select the appropriate folder and confirm the

selection.
Quality control
 The quality control process is initiated.

 The rainfall data file, containing days 8592of

data, is selected and opened for quality

control.
 The file is deemed acceptable for analysis.

Variables
 At first, the start file for rainfall data needs to

be selected.
 The date is given as default based on the

saved settings.
 The variable n1 needs to be set.

 A predictor variable must be selected from the

box file.
  The box file is accessed to select the ncp
folder.
 The first 11 predictor variables from the ncp

folder are selected.

 The total number of predictor values is 26, but

only the first 11 are selected initially.

 After selecting the 11 predictor variables, the

analysis can begin.

Correlation
 To select predictor variables, identify variables

with a p-value of zero or minimum and a

maximum partial r value.
 From a set of 11 data points, select two

predictor variables that meet the criteria of

minimum p-value (ideally zero) and maximum
partial r value.
 When encountering multiple variables with a

p-value of zero, prioritize the variables with

the highest partial r values.
Value
 Two values are discussed: the false value and

the large value, with the p-value being

minimum or zero and the partial value being
maximum
 The selection of data involves analyzing bank

and correlation, focusing on values where p is

0 or minimum
 Four predictor variables are selected where

the p-value is minimum or zero and the partial

r value is maximum
 The final two data points are selected based

on minimum p-value and maximum partial r

value
 The calibration model involves selecting a
rainfall dot file and an output folder for the par
file
 The model type is set to monthly, the process

is conditional, and the data period is selected

 Two final predictor variables from NCCP are

selected for calibration

 Calibration results are obtained, and the

process is completed
Scenario Generator
 The length should be changed to 'dcm 365'.

 The 'par' file from the 'calibrate model portion'

needs to be selected.
 The 'predictor variable folder' needs to be

selected from the 'box'.

Summary Statistics
 The output file from the scenario generator

should be selected as the input file.

 The output file will be a text format file

containing the predictor of the prediction file.

 The observed file and model file need to be

selected to compare the results.

Data Selection
 The user should input the desired date range

for future rainfall predictions.

 The input file path needs to be selected.

 Two predictor files, RCP 4.5 and RCP 8.5, are

available for predicting future rainfall.

Prediction
 After selecting the input file, the analysis

period is determined, and the analysis plot is

considered the most important part of the
prediction site.
 For prediction, input dates must be carefully
selected. For example, to predict for 2025, the
analysis start date and index are set based on
a 15-year data range from 2000 to 2050.
 The middle value of the data range is
identified, and the data is divided into seven
years before and after the middle year,
totaling 15 years.
 For predicting 2025, the selected range is from
2001 to 2024 .
 The output file is then selected, and the data
source is modeled and analyzed.
 To predict for future years, the same procedure
is followed. For example, for 2045, the
previous year is 2038, and the data is divided
similarly.
 The results are compared by selecting the
appropriate data points and copying them to
an Excel file for further analysis.
 The chart for rainfall data is created, showing
observed periods and predicted rainfall
conditions for 2025 and 2045.
 The chart indicates temperature trends, with
periods of increasing and decreasing
temperatures.
 The same procedure is applied for different
scenarios, such as RCP 8.5, by selecting the
appropriate input and output files and
generating the data.
 Any issues encountered during the process
can be addressed by seeking further
assistance.
 Future videos will discuss temperature data,
including maximum and minimum
temperatures.
 Section II
Rainfall downscaling through Time series analysis
software

Section III
(RAINFALL DOWNSCALING BY GOOGLE EARTH ENGINE
USING JAVASCRIPT )
CODE USED TO PREDICT RAINFALL