ESTIMATION OF SEDIMENT YIELD IN

CHALAKUDY RIVER

A PROJECT REPORT
submitted by

MANEESHA K SANKAR
TCR18CEWR11

to

the APJ Abdul Kalam Technological University

in partial fulfillment of the requirements for the award of the Degree

of

Master of Technology
in
Water Resources and Hydroinformatics

Department of Civil Engineering

Government Engineering College Trichur

Thrissur

JULY 2020
 DECLARATION

I undersigned hereby declare that the project report “ESTIMATION OF SEDIMENT YIELD IN

CHALAKUDY RIVER”, submitted for partial fulfillment of the requirements for the award of

degree of Master of Technology of the APJ Abdul Kalam Technological University, Kerala is a

bonafide work done by me under supervision of Dr. Sumam K S. This submission represents my

ideas in my own words and where ideas or words of others have been included; I have adequately

and accurately cited and referenced the original sources. I also declare that I have adhered to ethics

of academic honesty and integrity and have not misrepresented or fabricated any data or idea or fact

or source in my submission. I understand that any violation of the above will be a cause for

disciplinary action by the institute and/or the University and can also evoke penal action from the

sources which have thus not been properly cited or from whom proper permission has not been

obtained. This report has not been previously formed the basis for the award of any degree, diploma

or similar title of any other University.

THRISSUR

26 /07/2020 MANEESHA K SANKAR

 DEPARTMENT OF CIVIL ENGINEERING

GOVERMENT ENGINEERING COLLEGE TRICHUR

THRISSUR

CERTIFICATE

This is to certify that the report entitled ‘Estimation of sediment yield in Chalakudy

River’ submitted by Maneesha K Sankar to APJ Abdul Kalam Technological University in

partial fulfillment of the requirements for the award of the Degree of Master of Technology in
Water Resources and Hydroinformatics is a bonafide record of the project work carried out by
her under my guidance and supervision. This report in any form has not been submitted to any
other University or Institute for any purpose.

Dr. Sumam K S
Internal Supervisor
Associate Professor
Govt. Engineering College Trichur
Thrissur.

PG Coordinator HEAD OF THE DEPARTMENT

 ACKNOWLEDGEMENT

First and foremost, I thank God Almighty for always being with me and giving me strength

and confidence for completing this work.

I express my sincere gratitude to my guide Dr. Sumam K S, Associate Professor, Department

of Civil Engineering for her excellent guidance, moral support, encouragements and technical

ideas throughout the course of my project work.

I express my sincere gratitude to Dr. C V Lal, Head of Civil Engineering Department,

Government Engineering College, Thrissur, and Dr. N. Sajikumar, PG co-ordinator for their

suggestions during my presentations and for providing me with an environment to complete the

project successfully.

I would like to thank Smt. Indhu. S, Sub. Divisional Engineer, Central Water Commission

Bangalore for her kind support for collecting the data for my work

I extend my sincere gratitude to all the faculty members of the Civil Department who have

directly and indirectly helped me in completing this work.

I express my sincere thanks to all my friends and family for their encouragement and moral

support throughout the course of my project work.

Maneesha K Sankar

i
 ABSTRACT

Neural networks are a user-friendly alternative to physically complex models for the prediction

of soil erosion. The purpose of the study is to find out the best Artificial Neural network model

for predicting the sediment yield by considering different input parameters like discharge,

rainfall, flow velocity and temperature. The best ANN models were selected for forecasting the

sediment deposited in Chalakudy river basin in the year 2017 – 2018 and the results are

compared with the observed data. The input data’s for a period of 28 years (1990 to 2018) were

considered for this study. Models were developed for two periods 1990-1997 and 2010-2017.

Five years of data were used for training the model and the remaining two years of input data

were used for validating the model. The training was conducted using the Levengerg –

Marquardt algorithm. For selecting the best performing model, minimum value of RMSE and

NSE were taken as criteria for the evaluation of model. Discharge – Rainfall models gave the

best results for predicting the sediment yield. The selected model predicted the sediment

deposited in the Chalakudy River for the year 2017-2018 with 84.9% of accuracy.

ii
 TABLE OF CONTENTS

ACKNOWLEDGEMENT i
ABSTRACT ii
LIST OF TABLES v
LIST OF FIGURES vi
Chapter 1 INTRODUCTION 1
1.1 General 1
1.2 Sedimentation 2
1.3 Problems causes by sedimentation 3
1.4 Numerical Models 4
1.5 Need for the study 5
1.6 Objectives of the study 6
1.7 Organization of the thesis 6
Chapter 2 LITERATURE REVIEW 7
2.1 General 7
2.2 Artificial Neural Network models 7
2.3 SOIL WATER ASSESSMENT TOOL model 10
2.4 Summary 13
Chapter 3 METHODOLOGY 14
3.1 Softwares used 14
3.1.1 MATLAB R2014a 14
3.1.2 Microsoft Office Excel 2007 14
3.2 Artificial Neural Network (ANN) 15
3.2.1 Historical background of ANN 16
3.2.2 Advantages of ANN 16
3 .2.3 Architecture of Neural networks 17
3.2.4 Multilayer Perception 23
3.2.5 Neural Networks Training Algorithm 24
3.3 ANN using neural network fitting tool in MATLAB 25
3.4 Models developed using different input combinations 32
3.5 Model Evaluation Criteria 33

iii
 3.6 SUMMARY 34
Chapter 4 STUDY AREA AND DATA DESCRIPTION 35
4.1 Location and extent of study area 35
4.2 Description of data 36
4.3 Summary 37
Chapter 5 RESULTS AND DISCUSSION 38
5.1 General 38
5.2 Interrelations among the variables 38
5.3 Temperature analysis 41
5.4 Results of Artificial Neural Network models 44
5.5 Results of models developed by considering temperature variation 68
5.6 Comparison of computed sediment yield with observed results 70
5.7 Summary 72
Chapter 6 CONCLUSION 73
REFERENCES 75

iv
 LIST OF TABLES

Table 1 Input Combinations 32

Table 2 Data Description 37
Table 3 Results obtained for all models developed for predicting sediment yield 67
Table 4 New ANN models developed for new periods 69
Table 5 Results obtained for the new models 69
Table 6 Results of Sediment deposited in the year 2017-2018 71

v
 TABLE OF FIGURES

Figure 3.1 An example of a simple feed-forward networkz 17

Figure 3.2 An example of a deep learning network 18
Figure 3.3 Perceptrons 19
Figure 3.4 A typical layout of Multilayer Perception network 24
Figure 3.5 Neural network fitting tool 25
Figure 3.6 Neural Network Fitting Tool –Select data 26
Figure 3.7 Neural Network Fitting Tool – selecting percentages 27
Figure 3.8 Network architecture 28
Figure 3.9 Train the network 28
Figure 3.10 Neural network training 29
Figure 3.11 Regression plots 30
Figure 3.12 Plot Fit of the model 30
Figure 3.13 Neural Network results from saving tab 31
Figure 4.1 Map of Chalakudy River up to Arangali G&D Site of CWC 35
Figure 5.1 Discharge and Sediment variation with respect to time 39
Figure 5.2 Rainfall and Sediment variation with respect to time 39
Figure 5.3 Flow velocity and Sediment variation with respect to time 40
Figure 5.4 Temperature and Sediment variation with respect to time 40
Figure 5.5 Temperature variations for the period 1990-2018 41
Figure 5.6 Graph of Rainfall from the period 1990-2018 42
Figure 5.7 Rainfall and Temperature variation vs. Time 42
Figure 5.8 Discharge and Temperature variation vs time 43
Figure 5.9 Velocity and Temperature variation vs time 43
Figure 5.10 Neural network of Model 1 44
Figure 5.11 Graph obtained for the sediment for a training period of 1990-1997 45
Figure 5.12 Regression plot for Model 1 45
Figure 5.13 Error Histogram of Model 1 46
Figure 5.14 Plot fit of Model 1 46

vi
Figure 5.15 Graph obtained for the observed and predicted sediment for a training
47
period of 1996-1997
Figure 5.16 Neural network of Model 2 47
Figure 5.17 Graph obtained for the observed and predicted sediment for a training
48
period of 2010-2015
Figure 5.18 Graph obtained for the observed and predicted sediment for a testing
48
period of 2016-2017
Figure 5.19 Neural network of Model 3 49
Figure 5.20 Graph obtained for the observed and predicted sediment for a training
50
period of 1990-1995
Figure 5.21 Graph obtained for the observed and predicted sediment for a testing
50
period of 1996-1997
Figure 5.22 Neural network of Model 4 51
Figure 5.23 Graph obtained for the observed and predicted sediment for a training
52
period of 2010-2015
Figure 5.24 Graph obtained for the observed and predicted sediment for a testing
52
period of 2016-2017
Figure 5.25 Neural network of Model 5 53
Figure 5.26 Graph obtained for the observed and predicted sediment for a training
54
period of 1990-1995
Figure 5.27 Graph obtained for the observed and predicted sediment for a testing
54
period of 1996-1997
Figure 5.28 Neural network of Model 6 55
Figure 5.29 Graph obtained for the observed and predicted sediment for a training
56
period of 2010-2015
Figure 5.30 Graph obtained for the observed and predicted sediment for a testing
56
period of 2016-2017
Figure 5.31 Neural network of Model 7 57
Figure 5.32 Regression plot of model 7 58
Figure 5.33 Error histogram of the model 7 58
Figure 5.34 Graph obtained for the observed and predicted sediment for a training 59
period of 1990-1995

vii
Figure 5.35 Graph obtained for the observed and predicted sediment for a testing
59
period of 1996-1997
Figure 5.36 Neural network of Model 8 60
Figure 5.37 Regression plot of model 8 61
Figure 5.38 Error histogram of the model 8 61
Figure 5.39 Graph obtained for the observed and predicted sediment for a training
62
period of 2010-2015
Figure 5.40 Graph obtained for the observed and predicted sediment for a testing
62
period of 2016-2017
Figure 5.41 Neural network of Model 9 63
Figure 5.42 Graph obtained for the observed and predicted sediment for a training
64
period of 1990-1995
Figure 5.43 Graph obtained for the observed and predicted sediment for a testing
64
period of 1996-1997
Figure 5.44 Neural network of Model 10 65
Figure 5.45 Graph obtained for the observed and predicted sediment for a training
66
period of 2010-2015
Figure 5.46 Graph obtained for the observed and predicted sediment for a testing
66
period of 2016-2017
Figure 5.47 The observed and the predicted values of the sediment deposited in the
71
year 2017-2018

viii
 CHAPTER 1

INTRODUCTION

1.1 GENERAL

Sedimentation is one of the serious environmental issues that is widely recognized all

across the world. It has an impact on water quality, a reduction in reservoir capacity

and low agricultural productivity due to the removal of top nutrient soils. Sediment

outflow from the watershed depends upon the process through which the soil materials

are removed, transported and deposited by runoff and precipitation. Estimation of

suspended sediment is necessary, to deal with issues like the design of reservoirs or

dams and the transport of pollutants and sediments through streams and rivers. The

increasing effect of sediment loads causes an enormous burden on water

infrastructures due to the increasing operation cost, particularly in the case of water

extraction and treatment for community water supply. The interaction between soil

texture, geology, land use, topography and climate change seems to be increasing the

number of sediment loads produced and transported from the watershed. The

production and transportation of sediment in the watershed are deeply connected to

the synergistic influences of land use and climate. As a result of temperature and

climatic change, indirectly affects the rainfall which will influence the frequency and

magnitude of sediment flow. The changes in land use and soil management activities

have a greater effect on sediment flow compared to climate variability. However, the

extent to which the intensity and magnitude of sediment flows are affected by climate

change and land use remains a topic for discussion.

1
1.2 SEDIMENTATION

Erosion is the movement of soil, sediment and rock particles by wind, water and ice

resulting from the weathering of geological features. Sedimentation occurs when

eroded material settles down on the water column as the speed of flow decreases. The

sediments are transported from the higher to the lower level by the flow of water and

deposited in the river bed, banks and a floodplain. Estuaries are formed by the mixing

of water and ocean sediments for creating a complex sedimentary environment. Water

bodies and their estuary are complex ecosystems, the sediment starts move and

deposited under the right conditions. The change in the size, type and bed material of

the channels is a result of changes in flow and sediment loads. A channel is found to

be stable when its water flow and sediment movements are balanced over time. In

either of these changes, the channel shall alter its slope, depth, width, the pattern of

meander, bed structure and vegetation density.

Sedimentation is a mechanism in which soil particles are separated and transported by

running water and deposited as a layer of solid particles in water bodies. It is the

phenomenon that soil particle attains a certain velocity under the action of the

gravitational field. This translation velocity is known as the sedimentation of the

settling velocity. Sedimentation is a complicated phenomenon that varies with

sediment yield of the watershed, transport rate and mode of deposition. The

accumulation of sediments reduces the storage capacity and lifespan of reservoir.

Sedimentation remains one of the most serious issues to river ecosystems across the

world. A research was performed on the 145 rivers on the world with long-term

sediment records and results demonstrate that about 50 percent of rivers have a

statistically significant downward trend due to sedimentation (Walling and Fang,

2
2003). Sumi and Hirose (2009) reported that the global gross storage capacity of the

reservoir is approximately 600 km3 and the annual sedimentation rate of the reservoir

is approximately 31 km3 (0.52 per cent). This means that by the year 2100, the global

storage capacity of the reservoir would be reduced to 50% at this sedimentation rate.

1.3 PROBLEMS CAUSES BY SEDIMENTATION

Sedimentation is often caused by practices which directly interfere with rivers like

widening of flood control channels, building dams, bridges or crossings

and uncontrolled access to livestock and clearing riparian vegetation. Natural

vegetation plays a key role in stabilizing sediments loads. Living plants and large

wood logs, slow the flow of water and reduce the erosive strength while sediment

deposition can be maintained by root systems.

Once river banks are cleared the vegetation, it becomes more vulnerable to erosion.

Flows are more likely to change the course of rivers and streams and to create new

flood channels. If the topsoil is stripped from the flood plain then the productive

agricultural land gets damaged. The excess sedimentation will harm waterways

ecological health and environmental values which leads to the accumulation of coarse

and fine sediments in the slow-flow areas. Coarse sediments can fill deep permanent

river pools which are valued habitats for aquatic fauna and wildlife refuges in summer

and drought.

Sediment accumulations may cause upstream flooding, divert the flow to the adjacent

bank of the stream to neighbouring land which results in further erosion. The deep

permanent pools of rivers may be filled with sediments to destroy the aquatic life.

Sediments suspended in the water column minimize the light penetration and

photosynthesize the ability of the algae and other aquatic plants. The movement of

3
fine sediments are closely related to seasonal blue-green algae blooms. As a result,

natural food chains became disturbed by excess sediments. The abundance and

distribution of aquatic plants and animals and marine diversity will be declined due to

sedimentation.

Also, the river segment can shorten dams and reservoir lifetime. The sediments that

used to flow along with the relatively fast-moving river water are not deposited in the

reservoir when a river is dammed and a reservoir is created. It is because the river

water that flows into the reservoir moves too slowly to keep the sediments suspended

and the sediment settles to the reservoir floor. Reservoirs fill gradually with sand and

mud, which ultimately make them unusable for their intended purposes.

1.4 NUMERICAL MODELS

The direct calculation is the most accurate sediment load estimation process, but the

sampling of sediment concentration is costly and it can’t be carried out regularly. For

the estimation of sediment loads, data-driven models such as numerical models have

proven their effectiveness.

Many techniques have been made use of forecasting sediment yield models with high

accuracy and the research in this field is continuing. Statistical models based on the

time series have been used for hydrological forecasting and some of the statistical

models are a simple regression model, multiple regression models, and autoregressive

moving average models. The statistical approach uses classical methods to analyse

historical data to develop forecasting models for estimating sediment concentrations.

Nonetheless, these models are not meant to represent the non-linear processes

involved in the transformation of rainfall into sediment yield. Artificial intelligence

(AI) methods meet this need. Major AI methods that have wide range of applications

4
in water resources engineering are Genetic Algorithm, Artificial Neural Network, and

Fuzzy Logic.

Artificial neural networks are regression estimators which can estimate any

measurable function with a high degree of accuracy, so it is well suited for this study.

The development of artificial intelligence in the field of hydrology and water

resources simulation has generated several promising results. In the past few years,

numerical models can be used for modelling rainfall-runoff forecasting, prediction of

surge water levels and flash flood forecasting and the result shows convincing.

The artificial intelligence computational methods provide a great benefit over

traditional modelling, by handling a large number of noisy data from complex and

nonlinear systems whose relations are not fully understood. These techniques can be

effectively exploited for the construction of powerful intelligent systems.

1.5 NEED FOR THE STUDY

For any watershed, estimation of sediment becomes one of the important criteria for

water quality analysis. The outflow of sediment from the watershed depends upon the

detachment, transport and deposition of soil particles by discharge and rainfall. The

estimation of sediment yield is expected to resolve issues like the construction of

reservoirs and dams, sediment - pollutant transport in lakes and streams. Hence it is

necessary to have an appropriate method to analyse the amount of sediment deposited

in the river so that proper measures has to be taken to reduce the reservoir

sedimentation and protect water resources. Since Chalakudy is one of the most

affected areas in Kerala Flood 2018, it is taken as the study area.

5
1.6 OBJECTIVES OF THE STUDY

The objectives of the study are as follows:-

 To study the interrelation among the variables such as Rainfall, Discharge,

Flow velocity, Temperature and Sediment.

 To develop the best combination among rainfall, discharge, flow velocity and

temperature in ANN by predicting sediment yield for the watershed and

comparing its prediction performance with observed data.

 The best model will be considered to estimate the quantity of sediment

deposited in Chalakudi river basin in the year 2017- 2018. Thus to analyse the

accuracy of predicted values of the model.

1.7 ORGANIZATION OF THE THESIS

This thesis consists of six chapters. Chapter 1 gives a basic introduction to

sedimentation and various numerical methods. Chapter 2 gives the literature review on

the estimation of sediment yield using different numerical methods, includes artificial

neural networks and soil and water assessment tool (SWAT). The theoretical aspects

and methods to be followed to attain the final results are discussed in Chapter 3.

Chapter 4 gives a brief overview of the study area and the collected data. Chapter 5

gives the results and discussions. Chapter 6 gives the concluding remarks of the study

conducted.

6
 Chapter 2

LITERATURE REVIEW

2.1 GENERAL

Various works of literature regarding numerical methods such as Soil Water

Assessment Tool (SWAT) and Artificial Neural Network (ANN) used for estimating

the sediment yield are selected for review. The best suitable works of literature for

predicting the sediment yield are discussed below.

2.2 ARTIFICIAL NEURAL NETWORK MODELS

Nagy et al. (2002) estimated the suspended sediment concentration (SSC) in rivers by

developing an ANN model. Froude number, Reynolds number, stream width ratio and

the mobility number were used as an input parameter for estimating the suspended

sediment concentration. The performance of the ANN model compared with sediment

discharge formulas. It was found that artificial neural network models perform better

than the sediment discharge formula.

Bhattacharya et al. (2005) proposed geomorphology based (GANN) models and

non-geomorphology based (NGANN) models for the prediction of sediment yield in

the watershed of Banha, Jharkhand. These models were used to predict sediment loads

and compare its performance with one of the regression models. From the study, it can

be analysed that both the GANN and NGANN performed well in the feed-forward

ANN with backpropagation algorithm.

Raghuwanshi et al. (2006) proposed a multilayer perception neural network with the

backpropagation algorithm is adopted for a small watershed to estimate daily and

weekly basis sediment yield and runoff. 7 years of data (1991- 1997) were used for

7
developing the ANN model. For training the model, five-year data were used and the

model was tested by using the remaining two years of data. Results show that artificial

neural networks (ANN) models predict the sediment yield and runoff more accurately

than the regression-based models. The study demonstrates that the prediction

performance of the model can be improved by increasing the no. of hidden nodes and

input parameters.

Cigizoglu and Kisi (2006) developed a range dependent neural network model for the

estimation of suspended sediment. The study shows the range dependent neural

network was found to be superior to conventional ANN applications in which a single

network is trained on the entire dataset to estimate the suspended sediment. It was

seen that, both the low and high observed sediment values are closely approximated

by the RDNN.

Zhu et al. (2007) conducted a study in the Long Chuanjiang River in China,

developed an ANN model for simulating monthly suspended sediment flux. In the

study model, average rainfall, temperature, rainfall intensity and flow discharge used

as input for predicting the suspended sediments. When proper variables were

considered, ANN models predict the monthly suspended flux with fairly good

accuracy.

Cigizoglu, H. K. and Alp, M., (2007) study compares two ANNs models such as

feed-forward backpropagation and radial basis function models to evaluate the relation

between hydro meteorological variables and total daily suspended sediment loads. The

study was conducted on Juniata catchment in the United States. Rainfall, flow and

suspended sediment load are used for training the ANN models. Satisfactory results

were obtained when it is compared to the conventional multilinear regression.

8
Cigizoglu and Partal (2008) proposed the co-utilization of wavelets and neural

networks to estimate the suspended sediment load in rivers. Discrete wavelets

transform (DWT) and the new wavelet series uses the sum of combined wavelet

components as input for developing an ANN model. The predictions of Wavelet-

ANN model is found to be more superior to the conventional ANN and SRC models.

Mutlu et al. (2008) compared the simulation of stream flow at four stations using

multilayer perceptron and RBNN models. The flow prediction at multiple stations

gives a satisfactory performance by both the models. The multilayer perception model

was performed slightly better than the other model.

Agarwal et al. (2009) compared the performance of single and multi input linear

transfer function models and backpropagation artificial neural network (BPANN)

model for the estimation of runoff and sediment yield. The results show that the single

input linear transfer functions were more effective than the multi-input linear transfer

function (MI-LTF) and ANN models for forecasting runoff and sediment yield.

Wang et al. (2009) evaluated the performances of ANN models to Generalized

Regression Neural Networks models (GRNN) and the Back Propagation (BP) models.

From the study, it is found that ANN models give a better result than the regression

model for the prediction of sediment yield.

Rajaee et al. (2009) evaluated the performance of different models for simulating

Suspended Sediment Concentration. The study was conducted on Little Black River

and Salt River gauging stations in the United States. The daily river discharge and

SSC data were used for training the model. ANN model predicted the SSC values

more accurately than the MLR and SRC models. The findings also showed that the

ANN model was able to accurately estimate cumulative SSL.

9
Yenigün et al. (2010) developed ANNs models for the estimation of sediment yield

using feed forward and backpropagation algorithm. Results shows the effectiveness

and efficiency of this method are compared with classical methods. In this study, more

realistic values have been obtained with the ANN model compared with classical

equations. It can be concluded that when sediment measurement cannot be conducted

for a certain period, the accumulated sediment for the absent period may be estimated

by using the ANN technique without affecting the accuracy.

De Farias et al. (2010) developed an ANN model considering runoff and

meteorological data as input for estimating sediment yield. The study reveals that the

ANN model performs better than the multiple linear regression models and for

identifying the non-linear trends for variables ANN is the best the model.

Ajai Singh et al. (2013) developed two neural network models for the prediction of

monthly sediment yield on Nagwa watershed in Jharkhand. The standard

backpropagation (SBP) model and radial basis neural network (RBNN) model are the

two neural networks used for the study. The benchmark result shows that RBNN were

able to produce better results than SBP for the same data input. The accuracy of the

model can be improved by increasing the number of hidden nodes even though it is

time-consuming.

2.3 SOIL WATER ASSESSMENT TOOL MODEL

Kaur et al. (2003) developed a SWAT model by using a spatial decision support

system (SDSS) to estimate sediment yield and water quality in a large experimental

catchment in the Damodar-Barakar basin. Simulations of the total water and sediment

yields showed good to moderately good correlation coefficients. The study

10
demonstrated that the proposed SDSS could also be used to identify important areas

having high water and soil losses within the catchment.

Behera and Panda (2006) evaluated the SWAT applications and the study reveals

that the SWAT model simulates daily runoff, sediment yield and nutrient

concentration in runoff satisfactorily throughout the entire rainy season. The calibrated

model was successfully used for identifying the critical sub-watersheds and for the

development of best management practices.

Duan et al. (2009) developed a SWAT model for predicting sediment yields in the

Chaohe River upstream. For calibrating the model, the monthly runoff and sediment

yield data from the period 1985-1987 were used and 1988 to 1990 data were used for

validating the model. The model was calibrated and then validated to simulate

sediment yield. The NSE and R2 value obtained as 0.70 and 0.75 respectively. This

study shows that SWAT can be used to analyse erosion and watershed management in

the mountainous area.

Setegn et al. (2010) developed a SWAT model for the estimation of sediment yield in

the Anjeni gauged watershed in Ethiopia. Ten years of sediment yield data were used

for the study. The results show that the SWAT model succeeds in simulating sediment

yield within acceptable limits.

Oeurng et al. (2011) developed a SWAT hydrologic model for the prediction of a

sediment yields using a limited set of suspended sediment concentration data for the

period 2007−2009. The study was conducted in Coteaux Gascogne, France. The

model simulated the sediment yield of the basin within the acceptable limits, despite

the short span of observed suspended sediment data.

11
Prabhanjan et.al (2014) developed runoff and sediment yield model by integrating

soil and water assessment tool (SWAT) model and Geospatial techniques in

Khadakohol and Harsul watersheds in India. The monsoon months of a year were used

for calibrating the SWAT model and the other years were used for validating the

model. Due to the lack of availability of the observed data, the parameters for

estimating runoff and sediment were regionalized. The results obtained from the

experiment are promising.

Yesuf et al. (2015) simulated sediment yield using the SWAT model in the Maybar

gauged watershed of Ethiopia. 14 years of historical sediment data were used for

developing the model, which gives a satisfactory result.

Nagraj S. Patil and Priyanka (2016) developed the Soil and Water Assessment Tool

(SWAT) model and is used for the estimation of sediment yield for Malaprabha sub-

basin. SWAT uses the modified Universal Soil Loss Equation for the sediment yield

calculation. Before calibration, the observed values and stimulated peaks values

weren't matching with each other. After performing the number of trials, R2 and NSE

values reached the desired limit.

Based on the literature reviewed, SWAT models can be used for the prediction of

runoff and sediment yield but sometimes the results obtained between the observed

and the simulated values doesn't show a good match. If the performance of the

artificial neural network is considered, it gives better results even though if data are

scarce and is selected for the present study.

12
2.4 SUMMARY

In this chapter, different works of literature regarding various numerical methods used

for the estimation of sediment yield are discussed in the chronological order. Section

2.1 gives a brief introduction about the numerical tools used for the prediction of

sediment yield. Section 2.2 discussed the literature regarding the forecasting models in

detail by using Artificial Neural Network. From each study, merits and demerits of

models were discussed. In section 2.3 performance of SWAT models for the

prediction of sediment yield are discussed.

13
 CHAPTER 3

METHODOLOGY

3.1 SOFTWARES USED

The software’s used for this study are MATLAB R2014a and Microsoft Office Excel

2007. A brief description of the software's is given in the following subsections.

3.1.1 MATLAB R2014a

MATLAB - Matrix Laboratory is a multi-paradigm platform and fourth-generation

language for programming. MATLAB, a licensed programming language developed

by MathWorks enables users to manipulate the matrix, plotting functions and graphs,

implementing algorithms. It create an user interfaces to communicate with other

language programs such as C, C++, Java, Fortran and Python.

Although MATLAB is intended primarily for numerical computing, an optional

toolbox uses the MuPAD symbolic engine, allowing access to symbolic computing

capabilities. An additional package, Simulink adds graphical multi-domain simulation

and model-based design for dynamic and embedded systems.

For this study, MATLAB will be used for artificial neural network (ANN) analysis.

ANN models will be prepared using Neural Network toolbox. Neural Network Fitting

Tool GUI can be opened by using the command 'nftool’.

3.1.2 Microsoft Office Excel 2007

Microsoft Excel is a spreadsheet application developed by Microsoft for Microsoft

Windows, Mac OS X, and iOS. It features calculations, graphing tools, pivot tables

and a macro programming language is called Visual Basic for Applications. It has

14
been a very widely applied spreadsheet for these platforms, especially since version 5

in 1993. Excel forms part of Microsoft Office.

Microsoft Office Excel 2007 included in Office 2007. This release was a major

upgrade from the previous version. Similar to the other updated Office products, Excel

in 2007 used the new Ribbon menu system.

Added functionality included the SmartArt set of an editable business diagram. Also

added was improved management of named variable through the Name Manager, and

much-improved flexibility in formatting graph, which allow (x,y) coordinate labelling

and lines of the arbitrary weight. Several improvement two pivot tables were

introduced. Also like other Office products the Office open XML file format were

introduced include in .xlsm for a workbook with macros and .xlsm for a workbook

without macros.

3.2 ARTIFICIAL NEURAL NETWORK (ANN)

An artificial neural network (ANN) is an information processing paradigm inspired by

biological nervous systems, like the processing of information by the brain. The novel

structure of the information processing system is the key aspect of this model. It

consists of many highly interconnected processing elements (neurons) which work

together to solve specific problems. An artificial neural network learns by examples,

just the way human learns. An ANN is designed through a learning algorithm for a

particular task, such as pattern recognition or data classification. Learning in a

biological system involves the adaptation of the synaptic connections between the

neurons.

15
3.2.1 Historical background of ANN

The simulations of neural networks seem to be a recent innovation. However, before

computers came into use, this sector had survived at least one setback and several

eras.

In 1943, neurophysiologist Warren McCulloch and the logician Walter Pits created the

first artificial neuron. However, the technology at that time allowed them not to do

much.

3.2.2 Advantages of ANN

Neural networks can be used to extract patterns and to detect trends that are too

complex to be noticed through both human and other computer techniques with the

remarkable ability to drive meaning from complicated or imprecise data. A trained

neural network can be called an "expert" in the information category of research. This

expert can be used to make a prediction in new scenarios and answer questions of

"what if."

Other advantages include:

1. Adaptive learning: An ability to learn how to do tasks based on the data given

for training.

2. Self-organisation: An ANN can create its organisation or representation of the

information that receives during learning time.

3. Real-time operation: ANN computations may be carried out in parallel, and

special hardware devices are being designed and manufactured which take

advantage of this capability.

16
 4. Fault tolerance wire redundant information coding: Partial destruction of a

network leads to the corresponding degradation of performance. However,

some network capabilities may be retained even with major network damage.

3 .2.3 Architecture of Neural networks

Feed-forward networks

Feedforward ANNs allow signals to travel from input to output only one way. The

feedback loops do not have any effect on the output of the same layer. Feedforward is

usually simple networks associating input with outputs. They are commonly used to

recognize patterns. This type of organization is called the bottom-up or top-down

organization.

Feedback networks

The feedback network can have signals going in both directions through the network

loops. It is a very strong and complicated networking system. The feedback network is

a highly complex system which continually changes until it finds a balance. It stays in

the balance until their information updates or a new equilibrium is found.

Figure 3.1 An example of a simple feed-forward network

17
Network layers

The most common form of the artificial neural network consists of three unit layers.

An "input" unit layer is connected to hidden a unit layer which is then connected to an

"output" unit layer (refer figure 3.1). The input unit activity represents the raw

information that is transmitted to the network. The hidden unit's activity is determined

by the activities of the inputs and the weight of the connection between the input and

the hidden units.

Figure 3.2 An example of a deep learning network

The performance of the output unit depends on the operation of the weights and

hidden node. This simple form of network contains the hidden units in it, which helps

to construct their inputs layout freely. The weights between the input and the hidden

units are calculated by the activation function of the input unit and the hidden unit.

Single layer and multi layer architecture can be defined in perceptron. The single-layer

structure, with all elements linked, is the most common case and is more efficient than

18
multi-layer organisations. Units are often numerated by layers in multilayer networks

rather than adopting a standard numbering.

Perceptrons

The most influential work on neural networks in the '60s went under the heading of

“perceptrons" a term defined by Frank Rosenblatt. The perceptron (figure 3.3) turns

out to be a McCulloch-Pitts ‘MCP’ model (neurone with weighted inputs) with some

additional, fixed, pre-processing. Unit labelled X1, X2, Xm are called Association units

and their role is to extract from the input images which has unique and localized

features. Perceptrons emulate the fundamental idea behind the visual mammalian

system. They were primarily used to identify patterns, while their capabilities

increased a lot more.

Figure 3.3 Perceptrons

19
The learning process

Pattern memorisation and the subsequent network response can be divided into two

general paradigms:

1. In Associative mapping, the networks produce a particular pattern on the set of

input units whatever another particular pattern is applied to the set of input units.

The associative mapping can generally be broken down into two mechanisms:

i. Auto Association: An input pattern is connected to itself and the input and

output state coincides. This is used to complete the pattern and to generate a

pattern when a portion of it or a deformed pattern is introduced. In the second

case, the network stores pairs of patterns that link two pattern sets.

ii. Hetero Association is based on recall mechanisms:

 Nearest neighbour recall in which the output factors are generated

corresponding to the stored input pattern similar to the proposed pattern.

 In Interpolative recall, the output pattern is a dependent interpolation of the

stored patterns which corresponds to the proposed pattern.

Regularity detection: learns to respond to a particular property of the input patterns

which has a particular 'meaning' and this kind of learning process is important for

future exploration and representation of information. Whereas in the associative

mapping, the network stores the relationships among patterns.

Each neural network knows the weight of the connections. The information stored in

the network as function learning rule for changing the values of the weight and it can

distinguish between fixed networks and adaptive networks:

 Fixed networks in which weights can be changed, ie., . In such

networks, the weights are fixed as a prior according to the problem to solve.

20
  Adaptive networks which can change their weights, ie., ≠ 0.

All methods of learning for adaptive neural networks can be divided into supervised

learning and unsupervised learning:

Supervised learning involving an external teacher, such that each output unit is aware

of the response it gets on the input signal. Global information may be needed during

the learning process. Concepts of supervised learning comprise error correction,

reinforcement and stochastic learning. The important point in supervised learning is

the convergence of errors, which is the reduction of errors between the expected and

measured unit values. The goal is to identify a set of weights that minimizes the error.

The least medium square (LSM) convergence is a well-known approach common to

many learning paradigms.

Unsupervised learning does not use an external instructor and is based on local

knowledge only. It is also called self-organization and organizes information

presented to the network itself and detects its emerging collective properties.

Unsupervised learning paradigms are Hebbian and competitive learning. A neural

network can learn off-line when the process of learning and operation are distinct. A

neural network can learn digitally if it learns and operates simultaneously. Usually,

supervised learning is done off-line, whereas unsupervised learning is done on-line.

Transfer function

The action of an ANN (artificial neural network) depends on both the weight and the

input-output function specified for the units. Typically, this function comes under one

of three categories which are linear, threshold and sigmoid.

i. For Linear/Ramp units, the output activity is proportional to the total

weighted output.
21
 ii. For threshold units, the output is set at one of two levels, depending on

whether the total input is greater than or less than some threshold value.

iii. For sigmoid units, the output varies continuously but not linearly as the input

changes. Sigmoid units bear a greater resemblance to real neurones than the

linear or threshold units, but all these must be considered rough

approximations. To make a neural network performs some specific task, it

must choose how the units are connected and set the weights on the

connections appropriately. The connection determined whether it is possible

for one unit to influence another and the weights specify the strength of the

influence.

A three-layer network performs a particular task by using the following procedure:

1. Train the network with example which consists of a pattern of input unit

activities and the required pattern of output unit activities.

2. Then test how well the actual network output fits with the expected outcome.

3. The weight of each link adjusts so that the network correctly measures the

desired output.

The Back Propagation Algorithm

To minimize the error between the desired output and the actual output, the neural

network must be trained to adjust the weight of each variable. Through this process,

the neural network architecture to calculate the weight error derivative (EW). EW

computes the error changes for a slight difference in the weights. For determining EW,

backpropagation algorithm is the most widely used method.

In the backpropagation algorithm, the entire unit in the network is linear. The

algorithm computes each EW by first computing the EA, the rate at which the error

22
changes at the activity level of a unit is changed. For output units, the EA is simply the

difference between the actual and the desired output. To compute the EA for a hidden

unit in the layer just before the output layer, then identify all the weight between the

hidden units and the output units to which it is connected. Then multiply those weights

by the EA of those output units and add the products. This sum equals the EA for the

chosen hidden unit. After calculating all the EAs in the hidden layer just before the

output layer, compute the EAs for the other layers, moving from layer to layer in

opposite direction in the way activities propagate through the network. This is called

backpropagation. Once the EA has been computed for a unit, it is straightforward to

compute EW for each coming connection of the unit. The EW is the product of EA

and the activity through the incoming connection.

3.2.4 Multilayer Perception

One of the most popular ANN architecture is Multilayer Perception. Perceptron is a

simple neuron network. The perceptron forms a combination of linear relationships by

computing a single output from multiple real-value inputs based on input weights and

non-linear transfer function as shown schematically in Figure 3.4. Mathematically, the

MLP can be represented as:

(3.1)

Where wi represents the weight vector; pi is the input vector (i = 1....n); b is the bias; f

is the transfer function, and y is output.

23
 Figure 3.4 A typical layout of Multilayer Perception network [30]

3.2.5 Neural Networks Training Algorithm

The Levenberg - Marquardt algorithm is designed to approach the second-order

training speed without having to compute the Hessian Matrix. In most of the cases, the

best results for ANN models are given by the LM algorithm. So for the study, the LM

algorithm is selected. To minimise the global error E is the objective of the training

algorithm. Error E can be calculated by the following equation:

Where M is the total no. of training patterns,

Em is the error of the training pattern.

Em is derived as:

24
Where the total number of output nodes denoted by n,

is the network output at the kth output node and

target output at the kth output node.

3.3 ANN USING NEURAL NETWORK FITTING TOOL IN

MATLAB

1. Use ‘nftool’ command for opening the Neural Network Fitting Tool app.

Figure 3.5 Neural Network Fitting Tool

2. Click Next to proceed.

25
 Figure 3.6 Neural Network Fitting Tool –Select data

2. Load the Input and Target data saved from the MATLAB workspace.

3. Press the Next button to move to the validation and testing data window, shown

in the following figure given below.

26
 Figure 3.7 Neural Network Fitting Tool – selecting percentages

4. 15% of the original data are used for the validation and testing of data sets. Then

click Next.

5. In hidden layers, hidden neurons are set as 15. This value can be changed in

another run if the network doesn't perform as per the expectation.

27
 Figure 3.8 Network Architecture

8. Click Next.

Figure 3.9 Train the network

9. Click Train.

28
 Figure 3.10 Neural network training

10. From the Plots section, Click Regression. The regression plots of training,

validation, test and output data are shown in figure 3.11.

11. To know the network response, fit of the network can be plotted for single input/

output problems. To get fit diagram click Fit button under the plots pane.

29
 Figure 3.11 Regression Plots

Figure 3.12 Plot Fit of the model

30
12. To evaluate the network, Click Next in the Neural Network Fitting Tool.

If the performances of the network are not satisfactory then the network can be tested

with a new set of data. To improve the performance of the model following steps can

be done:

i. Train the model again.

ii. Increasing the number of neurons

iii. Train the model with a larger data set.

13. Once the performance of the networks is satisfactory then click Next.

14. The network is saved in the workspace as net1 and to generate simple script, click

on the simple script tab to reproduce all the results to solve similar kind of

problems.

15. When results are saved, Click Finish.

Figure 3.13 Neural Network results from saving tab

31
3.4 MODELS DEVELOPED USING DIFFERENT INPUT

COMBINATIONS

In this study, different combinations of four input variables such as Discharge,

Rainfall, Flow velocity and Temperature will be considered for training the ANN

model to forecast sediment yield. Models will be developed for two different periods

1990-1997 and 2010-2017. From these 14 models, the best model will be used for

forecasting the sediment yield.

Table 1 Input Combinations

Model Model
Model Period Input Layer Output layer
Name Parameters

M1 1990 -1997

M2 2010 – 2017 Q S Q S
t, t t t

M3 1990 -1997

M4 2010 – 2017 P S P S
t, t t t

M5 1990 -1997

M6 2010 – 2017 V S V S
t, t t t

M7 1990 -1997

M8 2010 – 2017 T , Q and S T , Qt S

t t t t t

M9 1990 -1997

M10 2010 – 2017 T , P and S T , Pt S

t t t t t

M11 1990 -1997

M12 2010 – 2017 P , Q and S P t , Qt S

t t t t

M13 1990 -1997

M14 2010 – 2017 P , V and S P ,V S

t t t t t t

32
Where Q = Discharge in m3/sec,
t

P = Precipitation in mm,
t

F = Flow velocity in m2/sec,

t

T = Temperature in ˚C
t

And S = Sediment Yield in g/litre.

t

3.5 MODEL EVALUATION CRITERIA

World Meteorological Organisation (WMO) and other researchers (Nash and

Sutcliffe) proposed different methods of verification criteria for the evaluation and the

comparison of the numerical models. For the present study, the root means square

error (RMSE) and the Nash Sutcliffe coefficient (NSE which is also called the

coefficient of simulation efficiency COE) is used as the numerical indicators. The

equation for calculating RMSE and COE value are given below,

Where, is the observed value of sediment;

is the predicted values sediment;

is the average value of observed sediment;

33
 and is the number of observations.

Model errors can be indicated by the value of the RMSE in terms of the dimensioned

quantity. A perfect match between the observed and predicted values indicates that

RMSE and COE a value are close to zero and 1 respectively.

3.6 SUMMARY

This chapter gives a brief introduction to the software used for the prediction of

sediment yield in section 3.1. In section 3.2, gives the basic concepts of Artificial

Neural Network and working methods. The methodology of Artificial Neural Network

GUI tool and the combination of different inputs used for the study are described in

section 3.3 and 3.4. In the last section of this chapter, the criteria used for evaluation

of the model are discussed.

34
 CHAPTER 4

STUDY AREA AND DATA DESCRIPTION

4.1 LOCATION AND EXTENT OF STUDY AREA

The Chalakudy River is one of the longest rivers of Kerala. The five streams

Parambikulam, Kuriarkutty, Sholayar, Karappara and Anakkayam originating

from the Anamalai Hills of Western Ghats, are confluence to form Chalakudy River.

The rivers Parambikulam and Sholayar originate in the Tamilnadu whereas the rivers

Karappara and Kuriarkuthti originate from the Palakkad district of Kerala. At 470m

above the Mean sea level, the Parambikulam joins with the Kuriarkutty River and the

river Sholayar is joined by 9 km down. The Karappara and Anakkayam meet the main

river about 455 m and 8 km further down at 365 meters above the MSL respectively.

The river reaches the plains of Athirapally by passing through the dense forests.

Figure 4.1 Map of Chalakudy River Up To Arangali G&D Site of CWC [29]

35
The Chalakudy River basin, in the South Western corner of peninsular India, lying

between 100 10’ - 100 35’ N latitudes and 760 15’ - 770 15' E longitudes represents a

central cross-section of the state of Kerala. The Chalakudy basin shares its eastern and

northern watershed boundary with Bharathapuzha basin, North-Western boundary

with Karuvannur basin and the southern boundary with Periyar basin. The basin

consists of about 30,000 hectares of wetlands. The annual rainfall in the basin is

around 3000 mm. The Chalakudy River has a total drainage area of 1704 Km2 and the

area of 1404 Km2 is in Kerala and the remaining area of 300 Km2 is in Tamilnadu.

The length of the river is about 130 kilometres. The Chalakudy river basin in the

ecological and political milieu of the Western Ghats and the state of Kerala and

provides an account of the resource development trajectory and dependences of the

hydrological basin.

4.2 DESCRIPTION OF DATA

Data used for the present study are collected from Arangali Hydrological observation

station, situated in Trichur district Kerala, under the Central Water Commission.

Central Water Commission has built a series of sediment observation stations and

stream gauging stations for observing hydrological parameters. Periyar River is the

main tributary in the hydrological observation station and its sub tributary is

Chalakudy River. The catchment area is 1342 Km2 and lies in the Longitude of

76°18’55” and Latitude of 10°16’53”.

The data required for this study are Discharge, Rainfall, Flow velocity, Temperature

and Sediment Yield. The data for the present study is described in the table given

below:

36
 Table 2 Data Description

Data Required Data Description

The daily discharge in m3/s from the year 1990 Jan to 2018
Discharge
Dec from Arangaly station.

The daily precipitation data from the year 1990 Jan to 2018
Rainfall
Dec from Arangaly station

The daily flow velocity data from the year 1990 Jan to 2018
Flow velocity
Dec from Arangaly station

The daily temperature from the year 1990 Jan to 2018 Dec
Temperature
from Arangaly metrological station

The daily sediment data from the year 1990 Jan to 2018 Dec
Sediment Yield
obtained from Arangaly station

4.3 SUMMARY

This chapter consists of two sections. Section 4.1 describes the location and extends of

the study area. The different data required for the project are also discussed.

37
 CHAPTER 5

RESULTS AND DISCUSSION

5.1 GENERAL

Throughout this analysis, the various combinations of discharge, precipitation, flow

velocity and temperature were used as input variables to develop monthly sediment

models. The description of different models to be developed is given in table 1. For

developing the models, seven years (1990-1997 and 2010-2017) of monthly data of

discharge, rainfall, flow velocity and temperature were used. For training the model

five years of data will be used and the remaining two years of data will be used for

testing the model. The trial and error method will be used for determining the number

of neurons in the hidden layers. The best ANN model will be selected based on COE,

R2 and RMSE values. The different process involved in the study will be discussed in

detail in the following sections.

5.2 INTERRELATIONS AMONG THE VARIABLES

For predicting the sediment yield, ANN models will be developed by considering

discharge, rainfall, flow velocity and temperature as an input. To understand the

relationship among the input variables with the sediment yield, different graphs are

plotted between the input variables such as discharge, rainfall, flow velocity,

temperature and the sediment yield.

A graph is plotted to relate the discharge and sediment yield with respect to time from

the year 1990 to2018 is shown in figure 5.1.

38
 400.00 0.0600

350.00
0.0500

DISCHARGE (m3/s)
300.00

SEDIMENT( g//l)
0.0400
250.00

200.00 0.0300

150.00 Discharge
0.0200
100.00 Sediment
0.0100
50.00

0.00 0.0000
1994

2010
1990
1991
1992
1993

1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009

2011
2012
2013
2014
2015
2016
2017
2018
YEARS

Figure 5.1 Discharge and Sediment variation with respect to time

From the graph, it is evident that the changes in the trend of discharge are similar to

the changes in the sediment yield except for a few years. Similarly, the graphs are

plotted between rainfall and sediment yield, flow velocity and sediment yield,

temperature and sediment yield which are given in the figures 5.2, figure 5.3 and

figure 5.4 respectively.

1400 0.0600

1200 0.0500

1000
SEDIMENT (g/l)

0.0400
RAINFALL ( mm)

800
0.0300
600 Rainfall
0.0200 Sediment
400

200 0.0100

0 0.0000
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018

YEARS

Figure 5.2 Rainfall and Sediment variation with respect to time

39
 3.5 0.0600

3
FLOW VELOCITY (m 2/s) 0.0500

2.5
0.0400

SEDIMENT (g/l)
2
0.0300
1.5

0.0200
1

0.0100 Flow Velocity

0.5
Sediment
0 0.0000
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
YEARS

Figure 5.3 Flow velocity and Sediment variation with respect to time

0.0600 40

35
0.0500
30
TEMPERATURE (˚C)
0.0400
SEDIMENT(g/l)

25

0.0300 20
Sediment
15
0.0200 Temperature

10
0.0100
5

0.0000 0
1990

1994

2002

2010

2018
1991
1992
1993

1995
1996
1997
1998
1999
2000
2001

2003
2004
2005
2006
2007
2008
2009

2011
2012
2013
2014
2015
2016
2017

YEARS

Figure 5.4 Temperature and Sediment variation with respect to time

By analysing all four graphs, it can be concluded that the changes in discharge,

rainfall and flow velocity are similar to the variation in the sediment yield. From the

graphs plotted between Discharge, Rainfall, flow velocity and sediment yield, it's clear

40
that these input variables shows a good relationship with the sediment yield so it can

be used as an input parameter for predicting the sediment yield. However, the

variation in the temperature is not similar to the variation of sediment yield. So it is

necessary to analyse more about temperature before considering it as an input

parameter for predicting the sediment yield.

5.3 TEMPERATURE ANALYSIS

For the further analysis of temperature, a graph of temperature vs time (Fig 5.5) is

plotted from the year 1990 -2018. For plotting the graph, the monthly average

temperature is considered.

40

35

30
TEMPERATURE in ˚C

25

20

15 Temperature

10

0
1994

2016
1990
1991
1992
1993

1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015

2017
2018

YEAR

Figure 5.5 Temperature variations for the period 1990-2018

By analysing the graph, one can observe a drop in the monthly average of temperature

from the year 2004 onwards till 2014 and the graph starts to rise from the year 2014.

This change in the pattern of average temperature is a significant variation, so it is

necessary to check whether this variation has any influence on the other input

parameters.

41
A graph of rainfall vs time is plotted to verify the influence of temperature on rainfall.

Figure 5.6 shows the plot of rainfall from the year 1990-2018. From the graph, it was

found that rainfall shows variation throughout the years.

1400
1200
RAINFALL in mm

1000
800
600
rainfall
400
200
0
1993

2010
1990
1991
1992

1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009

2011
2012
2013
2014
2015
2016
2017
2018
YEARS

Figure 5.6 Graph of Rainfall from the period 1990-2018

To understand whether the temperature variation has any influence on the

precipitation variation, a graph of rainfall and temperature vs time is plotted. From the

graph, it is obvious that the variation of temperature does not affect the rainfall

change.

1400 40

1200 35
RAINFALL in mm

30
1000
TEMPERATURE

25
800
20
600
15
400 10
Rainfall
200 5
Temperature
0 0
1991

2008
1990
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007

2009
2010
2011
2012
2013
2014
2015
2016
2017
2018

YEARS

Figure 5.7 Rainfall and Temperature variation vs. time

42
Similarly, graphs are plotted between discharge and flow velocity with temperature

which is shown in figure 5.8 and 5.9 respectively. It is evident from the graphs that

temperature changes do not have an impact on discharge and flow velocity as well.

400.00 40

350.00 35
DISCHARGE in m3/s

300.00 30

TEMPERATURE
250.00 25

200.00 20
Discharge
150.00 15
Temperature
100.00 10

50.00 5

0.00 0
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
YEARS

Figure 5.8 Discharge and Temperature variation vs time

3.5 40

3 35
FLOW VELOCITY in m2/s

30
2.5
TEMPERATURE

25
2
20
Flow
1.5
15 velocity

1
10
0.5 5

0 0
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018

YEARS

Figure 5.9 Velocity and Temperature variation vs time

43
By evaluating all four graphs, it can be concluded that the temperature has no

influence on rainfall, discharge, flow velocity and sediment, so that the temperature

can not be used as an input parameter for predicting the sediment yield.

5.4 RESULTS OF ARTIFICIAL NEURAL NETWORK MODELS

In this study, different combinations of rainfall, flow velocity and discharge

considering as input to develop monthly sediment yield prediction model. The

overview of the different models to be developed is given in Section 3.4 of Table 1.

The ANN models will be trained in the Levenberg Marquardt backpropagation

algorithm to achieve 0.01 root mean square error.

I. Model 1

In Model 1, discharge is taken as the input parameter for the period of 1990 – 1997

and sediment yield is taken as a target for developing an ANN model. Five-year data

will be used for training the model and the remaining two-year data will be used as

testing the model. Results obtained from model 1 are shown below

Figure 5.10 Neural network of Model 1

44
 0.0600
0.0500
SEDIMENT in g/l 0.0400
0.0300
0.0200
Observed
0.0100
Predicted
0.0000
June

June

June

June

June
March

March

March

March

March
September
December

September
December

September
December

September
December

September
December
1990 1991 1992 1993 1994 1995
YEARS

Figure 5.11 Graph obtained for the sediment for a training period of 1990-1997

Figure 5.12 Regression plot for Model 1

45
 Figure 5.13 and 5.14 Error Histogram and plot fit of Model 1

The RMSE value and COE obtained for the training the model as:

Model 1 RMSE COE

Training 0.0027 0.934427

The observed and predicted values shows a perfect match when RMSE equal to zero.

Similarly, COE value closer to 1 show that the observed and predicted values are in

good agreement with each other.

This trained model will be used for simulating the sediment yield for a period of 1996

-1997. The graph of observed vs predicted sediment is plotted again as in figure 5.15.

RMSE and COE value will be calculated for the testing model, to check its

compatibility with the trained model values.

RMSE and COE values obtained for testing the model as

Model 1 RMSE COE

Testing 0.002859 0.853889

46
 0.0250

SEDIMENT in g/l
0.0200

0.0150

0.0100

0.0050 Observed
Predicted
0.0000

February

February
June

June
October

April

October

April
December

December
August

August
1995 1996 1997
YEARS

Figure 5.15 Graph obtained for the observed and predicted sediment for a training

period of 1996-1997

RMSE value and COE value are calculated for the testing period, shows a slight

difference in the value of RMSE and COE for the training period. Still, results hold

well, because RMSE value and COE value are close to the standard result.

II. Model 2

In Model 2, discharge is taken as the input parameter for the period 2010-2017 and the

sediment yield is taken as the target for the development of the ANN model. Five-year

data (2010-2015) are used for training the model and the remaining two-year data are

used for testing the model. Results from Model 2 are shown below.

Figure 5.16 Neural network of Model 2

47
 0.0600
0.0500

SEDIMENT in g/l 0.0400

0.0300
0.0200
Observed
0.0100
Predicted
0.0000 May

May

May

May

May

May
January

January

January

January

January

January
September
September

September

September

September

September
-0.0100

2010 2011 2012 2013 2014 2015

YEARS

Figure 5.17 Graph obtained for the observed and predicted sediment for a training

period of 2010-2015

0.0250

0.0200
SEDIMENT in g/l

0.0150

0.0100
Observed
0.0050
Predicted
0.0000
May

July

May

July
January
February

January
February
June

June
March

March
October

October
April

April
November

November
September
September

December

December
August

August

2016 2017
YEARS

Figure 5.18 Graph obtained for the observed and predicted sediment for a testing

period of 2016-2017

The number of neurons in the hidden layer was determined by the trial-and-error

method. The RMSE value obtained for training and testing the model is 0.004 and

0.003 respectively which is close to zero value and COE value is obtained as 0.775

48
and 0.703 for training and testing the model which is nearer to the standard value 1.

Observed value and predicted value shows a good match with each other. By

comparing the results of 1990-1997 and 2010-2017 models, Model 1 gives a better

performance than model 2. From these two models, it can be concluded that discharge

can be used as an input parameter for predicting the sediment yield.

III. Model 3

In Model 3, rainfall is used as an input parameter for the period 1990-1997 and

sediment yield is used as a target for the development of the ANN model. Seventy

percentage data will be used to train the model, and the remaining thirty percentage

data will be used to test the model. Results from model 3 are shown below.

Figure 5.19 Neural network of Model 3

49
 0.0600

0.0500

0.0400
SEDIMENT

0.0300

0.0200 Observed

0.0100 Predicted

0.0000
June

June

June

June

June
March

March

March

March

March
December

December
September

September
December

September

September
December

September
December
1990 1991 1992 1993 1994 1995

PERIOD

Figure 5.20 Graph obtained for the observed and predicted sediment for a training

period of 1990-1995

0.0300

0.0250

0.0200
SEDIMENT in g/l

0.0150

0.0100
Observed
0.0050 Predicted

0.0000
July

May

July

May
January
February

January
February
June

June
March

October

March
October

April

April
November

November
September

December

September

December
August

August

1996 1997
YEARS

Figure 5.21 Graph obtained for the observed and predicted sediment for a testing

period of 1996-1997
50
The RMSE and COE values obtained for training the model is 0.0041 and 0.847

respectively and for testing the model it is obtained as 0.0033 and 0.8087. There is a

slight difference in the values of training and testing. Since the RMSE and COE

values are close to the standard values, one can ensure a good result.

IV. Model 4

In Model 4, Rainfall is taken as the input parameter for the period 2010-2017 and the

sediment yield is taken as the target for developing the ANN model. Five-year data

(2010-2015) are used for training the model and for the validation of the model, the

sediment yield is simulated for the year 2016 to 2017 and its performance is compared

with the observed data. Results from the Model 4 are shown as

Figure 5.22 Neural network of Model 4

51
 0.0600

0.0500

0.0400

0.0300
SEDIMENT in g/l

Observed
0.0200
Predicted
0.0100

0.0000

July
July

July

July

July

July
January

January

January

January

January

January
October

October

October

October

October

October
April

April

April

April

April

April
2010 2011 2012 2013 2014 2015
YEARS

Figure 5.23 Graph obtained for the observed and predicted sediment for a training

period of 2010-2015

0.0250

0.0200
SEDIMENT in g/l

0.0150

0.0100
Observed
0.0050 Predicted

0.0000
May

July

May

July
February
January
February

January
March

March
October

October
April

April
June

June
September

December

September

December
August

August
November

November

2016 2017
YEARS

Figure 5.24 Graph obtained for the observed and predicted sediment for a testing

period of 2016-2017

52
The numbers of neurons in the hidden layer were determined by the trial-and-error

method. Model performance can be improved by increasing the number of hidden

neurons. The RMSE value obtained for training and testing the model is 0.005234

and 0.003084 respectively which is close to zero and COE value obtained as 0.725553

and 0.672167 for training and testing the model which is nearer to standard value 1.

Observed value and predicted value shows a good match with each other. By

comparing the results of 1990-1997 and 2010-2017 models, Model 3 gives a better

performance than the model 4. From these two models, it can be concluded that

rainfall can be used as an input parameter for predicting the sediment yield.

Considering the discharge and rainfall models, discharge models perform more

efficiently than the rainfall models.

V. Model 5

ANN model 5 developed by considering flow velocity as the input parameter and

sediment yield as the target for a period of 1990-1997. Monthly average flow velocity

is used for developing the model. The data for the period from 1990 to 1995 is used

for training the model and the data for the period from 1996-1997 is used for

simulating sediment yield and its performance is compared with the observed data.

The results obtained from this model are given below.

Figure 5.25 Neural network of Model 5

53
 0.0600

0.0500

0.0400
SEDIMENT in g/l

0.0300

0.0200
Observed
0.0100 Predicted
0.0000
June

June

June

June

June
March

March

March

March

March
September
December

September
December

September
December

September
December

September
December
1990 1991 1992 1993 1994 1995
YEARS

Figure 5.26 Graph obtained for the observed and predicted sediment for a training

period of 1990-1995

0.0300

0.0250
SEDIMENT in g/l

0.0200

0.0150

0.0100
Observed
0.0050 Predicted

0.0000
February

February
October

October
April

April
June
June

December

December
August

August

1995 1996 1997

YEARS

Figure 5.27 Graph obtained for the observed and predicted sediment for a testing

period of 1996-1997

54
The RMSE and COE values obtained for training the model is 0.003596 and 0.88757

respectively and for testing the model it is obtained as 0.003418 and 0.791183. A

slight difference is observed in the values of training and testing. Since the RMSE and

COE values are close to the standard values, the results hold well. Considering the

models for the period 1990-1997, discharge model performs best followed by flow

velocity model and then by the rainfall model.

VI. Model 6

In Model 6, the flow velocity is taken as the input parameter for the period 2010-2017

and the sediment yield is taken as the target for developing the ANN model. Five-year

data (2010-2015) is used for model training and the validation of the model, the

sediment yield is simulated for the year 2016 to 2017 and its performance is compared

with the observed data. Results from Model 6 are shown below.

Figure 5.28 Neural network of Model 6

55
 0.0600

0.0500
SEDIMENT in g/l
0.0400

0.0300

0.0200 Oberved

0.0100 Predicted

0.0000
July

July

July

July

July

July
January

January

January

January

January

January
October
October

October

October

October

October
April

April

April

April

April

April
2010 2011 2012 2013 2014 2015
YEARS

Figure 5.29 Graph obtained for the observed and predicted sediment for a training

period of 2010-2015

0.0300

0.0250
SEDIMENT in g/l

0.0200

0.0150

0.0100
Observed
0.0050
Predicted
0.0000
May

July

May

July
January
February

January
February
March

June

March

June
October

October
April

April
November

September

November
September

December

December
August

August

2016 2017
YEARS

Figure 5.30 Graph obtained for the observed and predicted sediment for a testing

period of 2016-2017

56
The RMSE value is obtained for training and testing the model as 0.006156 and

0.003503 respectively which is close to zero whereas COE value obtained as 0.620358

and 0.577174 for training and testing the model which is not so close to the standard

value 1. In this model predicted value doesn’t show a good match with the observed

value. So this model can’t be used for predicting the sediment yield comparing its

performance with the other two models. By comparing the results of 1990-1997 and

2010-2017 models taking flow velocity as input, Model 5 gives a better performance

than the model 6. If discharge, rainfall, and flow velocity models were considered,

discharge models perform more efficiently than the rainfall models and flow velocity

models.

VII. Model 7

In Model 7, discharge and rainfall are taken as the input parameter for the period of

1990 – 1997 and sediment yield is taken as a target for developing an ANN model.

Here two input parameters were considered for predicting the sediment yield. Five-

year data, which is seventy percentages of data, is used for training the model and the

remaining two-year data is used for testing the model. Results obtained from model 7

are shown below.

Figure 5.31 Neural network of Model 7

57
 Figure 5.32 Regression plot of model 7

Figure 5.33 Error histogram of the model 7

58
 0.06

0.05

0.04

0.03
SEDIMENT in g/litre

0.02 Observed
Predicted
0.01

0
June

June

June

June

June
March

March

March

March

March
December
September
December

September

September
December

September
December

September
December
-0.01

1990 1991 1992 1993 1994 1995

YEARS

Figure 5.34 Graph obtained for the observed and predicted sediment for a training

period of 1990-1995

0.025

0.02

0.015
SEDIMENT in g/litre

0.01
Observed
0.005 Predicted

0
July

May

July

May
February
January
February

January
March

March
October

October
April

April
June

June
September

December

September

December
August

August
November

November

1996 1997

YEARS

Figure 5.35 Graph obtained for the observed and predicted sediment for a testing

period of 1996-1997

59
 The RMSE and COE values obtained for training the model is 0.0025708 and 0.9425

respectively and for testing the model it is obtained as 0.0026537 and 0.8744. There is

a slight difference in the values of training and testing. Since the RMSE and COE

values are close to the standard values, one can say that model holds a good result.

This model gives a better RMSE and COE values than discharge and rainfall, by

considering the inputs individually. Two input parameters give a better result than the

single input parameter models. So it’s evident that by increasing the input parameters

can improve the efficiency of the sediment yield prediction yield model.

VIII. Model 8

This model is developed by considering two input parameters such as discharge and

rainfall for predicting the sediment yield in the year 2010-2017. Five-year data is

considered for training the model and two-year data is used for testing the model and

its performance is compared with the observed data. Outputs obtained for this model

are plotted below.

Figure 5.36 Neural network of Model 8

60
 Figure 5.37 Regression plot of model 8

Figure 5.38 Error histogram of the model 8

61
 0.0600

0.0500
SEDIMENT in g/litre
0.0400

0.0300

0.0200
Observed
0.0100 Predicted

0.0000
May

May

May

May

May

May
January

January

January

January

January

January
September

September

September

September

September

September
-0.0100

2010 2011 2012 2013 2014 2015

YEARS

Figure 5.39 Graph obtained for the observed and predicted sediment for a training

period of 2010-2015

0.025

0.02
SEDIMENT in g/litre

0.015

0.01
Observed
0.005 Predicted

0
February

February
Month

October

October
April

April
June

June
December

December
August

August

Year 2016 2017

YEARS

Figure 5.40 Graph obtained for the observed and predicted sediment for a testing

period of 2016-2017

62
The RMSE and COE values obtained for training the model as 0.004 and 0.82733

respectively and for testing the model it is obtained as 0.003 and 0.75917. Since the

RMSE and COE values are close to the standard values, can say results hold good.

This model gives a better RMSE and COE values than discharge and rainfall, where

inputs considered as individually for 2010-2017 periods. Here also, two input

parameters give a better result than the single input parameter models. So it’s evident

that by increasing the input parameters can improve the efficiency of the sediment

yield prediction yield model.

IX. Model 9

In Model 9, flow velocity and rainfall are taken as the input parameters for a period of

1990 – 1997 and sediment yield is taken as a target for developing an ANN model.

Here two input parameter are considered for predicting the sediment yield. Five-year

data is used for training the model and the remaining two-year data is used for testing

the model. Results obtained from model 9 are shown below.

Figure 5.41 Neural network of Model 9

63
 0.06

0.05

0.04

0.03
SEDIEMNT in g/l

0.02
Observed
0.01
Predicted
0
March

March

March

March

March
June
June

June
September

June

June
September
December

December

September
December

September
December

September
December
1990 1991 1992 1993 1994 1995

YEARS

Figure 5.42 Graph obtained for the observed and predicted sediment for a training

period of 1990-1995

0.03

0.025
SEDIMENT in g/l

0.02

0.015

0.01
Observed
0.005
Predicted
0
July

May

July

May
January
February

January
February
March

March
October

October
April

April
June

June
September

December

September

December
August

August
November

November

1995 1996 1997

YEAR

Figure 5.43 Graph obtained for the observed and predicted sediment for a testing

period of 1996-1997

64
The RMSE and COE values obtained for training the model is 0.00302 and 0.92071

respectively and for testing the model it is obtained as 0.00303 and 0.83639. Since the

RMSE and COE values are close to the standard values, can say results hold a good.

This model gives better RMSE and COE values than by considering flow velocity and

rainfall as an individual input. Here also, two input parameters give a better result than

the single input parameter models.

X. Model 10

Model 10 is developed by considering two input parameters such as flow velocity and

rainfall for predicting the sediment yield for the year 2010-2017. Five-year data is

considered for training the model and two-year data is used for testing the model and

its performance is compared with the observed data. Outputs obtained for this model

are plotted as

Figure 5.44 Neural network of Model 10

65
 0.0600

0.0500
SEDIMENT in g/l
0.0400

0.0300

0.0200 Observed
Predicted
0.0100

0.0000

July
July

July

July

July

July
January

January

January

January

January

January
October

October

October

October

October

October
April

April

April

April

April

April
2010 2011 2012 2013 2014 2015

YEARS

Figure 5.45 Graph obtained for the observed and predicted sediment for a training

period of 2010-2015

0.0250

0.0200
SEDIMENT in g/l

0.0150

0.0100
Observed
0.0050 Predicted

0.0000
May

May
July

July
January

January
February

February
March

March
April

October

April

October
June

June

September
September

December

December
August

August
November

November

2016 2017

YEARS

Figure 5.46 Graph obtained for the observed and predicted sediment for a testing

period of 2016-2017

66
The RMSE and COE values obtained for training the model is 0.0048 and 0.7672

respectively and for testing the model it is obtained as 0.0030 and 0.6945. Since the

RMSE and COE values are close to the standard values, can say results hold a good.

Comparing the results obtained for the period 2010-2017, this model gives a better

RMSE and COE values than the flow velocity and rainfall where these inputs are

considered individually. Here also, two input parameters give a better result than the

single input parameter models. Comparing the performance of discharge- rainfall

model with flow velocity – rainfall models, discharge-rainfall models performs

slightly better than the flow velocity- rainfall model.

The overall output obtained for these 10 models are given in Table 3.

Table 3 Results obtained for all models developed for predicting sediment yield

RMSE COE
Model Period Input
Training Testing Training Testing

Model 1 1990-1997 0.003 0.003 0.934 0.854

Discharge
Model 2 2010-2017 0.005 0.003 0.771 0.703

Model 3 1990-1997 0.004 0.003 0.847 0.809

Rainfall
Model 4 2010-2017 0.005 0.003 0.726 0.672

Model 5 1990-1997 0.004 0.003 0.888 0.791

Flow Velocity
Model 6 2010-2017 0.006 0.004 0.620 0.577

Model 7 1990-1997 Discharge & 0.003 0.003 0.943 0.874

Model 8 2010-2017 Rainfall 0.004 0.003 0.827 0.759

Model 9 1990-1997 Flow Velocity & 0.003 0.003 0.921 0.836

Model 10 2010-2017 Rainfall 0.005 0.003 0.767 0.694

67
By analysing all these 10 models, it can be concluded that ANN models can be used

for the prediction of sediment yield. Discharge, rainfall and flow velocity can be used

as input for developing sediment yield. From the models, it is evident that models with

discharge as input gives values more close to observed values of sediment yield as

compared to those predicted using the rainfall and flow velocity as input.

Models developed for 1990-1997 performs better than 2010-2017 models. The

models with two input parameters performed better than the model with the single

input parameter. Out of 10 models, Model 7 gives the best result to predict sediment

yield and this model’s input is a combination of discharge and rainfall.

5.5 RESULTS OF MODELS DEVELOPED BY CONSIDERING

TEMPERATURE VARIATION

New models were developed by considering the change in the ANN model period. To

check whether the change in the period could improve the performance of ANN

models to predict sediment yield. New periods were selected based on temperature

variation. The ANN models will, therefore, be developed for the two new periods

1990-2004 and 2005-2014. Discharge and rainfall are input variables that provide the

best results of the ANN model for predicting sediment yield, so new models will be

developed by taking discharge and rainfall as input parameters. The details of the

models to be developed are shown in Table 4.

68
 Table 4 New ANN models developed for new periods

Model Output layer

Model Period Input Layer
Name

M11 1990 -2004

Qt St

M12 2005 - 2014

M13 1990 -2004

Pt St

M14 2005 - 2014

M15 1990 -2004

Pt, Qt St

M16 2005 - 2014

The outputs obtained for the new models are given in the table below. The table

compares the performance of models developed with the previous periods and new

periods.

Table 5 Results obtained for the new models

Model Period Input RMSE COE

1990-1997 0.003 0.934
Model 11 Discharge
1990-2004 0.004 0.812
2010-2017 0.004 0.770
Model 12 Discharge
2005-2014 0.005 0.796
1990-1997 0.004 0.847
Model 13 Rainfall
1990-2004 0.005 0.734
2010-2017 0.004 0.725
Model 14 Rainfall
2005-2014 0.006 0.676

1990-1997 0.003 0.943

Model 15 Discharge and Rainfall
1990-2004 0.002 0.910

2010-2017 0.004 0.827

Model 16 Discharge and Rainfall
2005-2014 0.005 0.807

69
Evaluating the performance of Model 11, gives a very good RMSE and COE value

that is similar to the standard values but comparing these values with the 1990-1997

Model, 1990-97 gives better results than the 1990-2004 Model. In the same way, all 6

models are reviewed, except model 12; all other models have lower performance than

the previous period models. By analyzing the table 5, it’s clear that models perform

better from 1990 to 1997 and 2010-2017 than the period corresponding to the

variation of temperature (1990-2004 & 2005-2014). Models developed considering

the variation of temperature doesn't improve the performance of the sediment yield

models.

From the study, it can be inferred that temperature is an independent parameter and

has no effect on sediment yield. Therefore, the temperature cannot be used for

modelling the sediment yield. ANN models developed over the period, considering

the variation in average temperature; do not improve the performance of the model.

5.6 COMPARISON OF COMPUTED SEDIMENT YIELD WITH

OBSERVED RESULTS

From section 5.4, Model 7 and Model 8 provide the best results for predicting

sediment yield. Sediment deposited in the years 2017 and 2018 are simulated using

Model 7 and Model 8, to test the efficiency of the model with the observed data.

Chalakudy is one of the most affected areas by the Kerala flood of 2018. To

determine the impact of the flood on sediment deposit in the Chalakkudy river, one

has to quantify it for the year 2017-2018.

The annual sediment yield for the years 2017 and 2018 is estimated. The difference

between annual sediment yields is the volume of sediment deposited in the year 2017-

70
2018. The observed sediment deposited value is compared to the predicted sediment

deposited using Model 7 and Model 8, to evaluate the performance of the ANN

models. Table 6 gives the results of sediment deposited in the chalakudy basin in the

year 2017-2018.

Table 6 Results of Sediment deposited in the year 2017-2018

Predicted using Predicted using

Observed Model 7 Model 8
(g/l) (1990-1997) (2010-2017)
g/l g/l

Sediment Deposited
0.142789 0.121282 0.1669
in the year 2017-2018

Sediment Deposited
0.180000
0.1669
0.160000
0.140000 0.142789
0.120000 0.121282
SEDIMENT (g/l)

0.100000
0.080000 Sediment Deposited
0.060000
0.040000
0.020000
0.000000
Observed (g/l) Model (1990-1997)Model (2010-2017)

Figure 5.47 The observed and the predicted values of the sediment deposited in the

year 2017-2018

From the graph (figure 5.45), it has to be pointed out that the sediment deposit

predicted by the model (1990-'97) and model (2010-'17) are close to the observed

71
value. Model 7 predicted with 84.94% of accuracy whereas model 8 predicted with

83.11% of accuracy.

5.7 SUMMARY

In this chapter, results and discussion of each model developed for this project is

explained in detail. Section 5.1 gives a general introduction to the different process

involved in the study. In section 5.2 discuss the interrelationship between input

variables and sediment yield, followed by a temperature analysis in section 5.3. A

detailed description of the ANN models and their results are shown in section 5.4.

Results and discussions of models developed by considering temperature variation are

explained in sections 5.5. Finally, section 5.6 provided a comparison of the computed

sediment yield with the observed results.

72
 CHAPTER 6

CONCLUSION

Artificial neural network models were developed to predict monthly sediment yield

from easily available information on discharge, precipitation and flow velocity.

Different neural network architectures were developed by considering the input

variables and hidden neurons. For developing the models seven years of data of

discharge, rainfall, flow velocity and sediment yield from the year 1990-2018 were

used. Models are developed for two periods 1990-1997 and 2010-2017. For training,

the model five years of data were used and remaining two-year data were used to test

the models. The conclusions arrived from the study are as follows.

 ANN models can be used for predicting sediment yield.

 Rainfall, Discharge and flow velocity can be used as an input parameter for

predicting the sediment yield.

 By increasing the input parameters and hidden neurons, can increase the

efficiency of the model to predict sediment yield.

 Combination of Discharge and Rainfall model gives the best result for

predicting the sediment yield.

 From the temperature analysis, it is concluded that the temperature is an

independent parameter and it has no effect on sediment yield. So temperature

can’t be taken for the modelling sediment yield which is contradictory to the

concept (N. S. Raghuwanshi (2006) [18].

73
  To forecast the sediment yield for any year, ANN models can be used if

rainfall and discharge are known.

Future scope of the study

It is necessary to have proper methods for predicting sediment yield at the basin so

that effective measures can be taken to reduce reservoir sedimentation and protect

water resources. Climate change is one of the biggest threats to global security and it

can be characterized by increasing temperature. Climate change has an indirect effect

on precipitation and surface hydrology, so it is necessary to study the effect of climate

change on sediment loads.

74
 REFERENCES

1. Agarwal, A., Rai, R. K., and Upadhyay, A. (2009) Forecasting of runoff and

sediment yield using artificial neural networks, Journal of Water Resources

and Protection, 1(05), 368–375.

2. Ajai Singh, Mohd. Imtiyaz, R. K. Isaac, and Denis, D. M. (2013)

Comparison of Artificial Neural Network Models for Sediment Yield

Prediction at Single Gauging Station of Watershed in Eastern India, Journal of

Hydrologic Engineering, 18(1): 115-120.

3. ASCE Task Committee on Application of Artificial Neural Networks in

Hydrology. 2000a. Artificial neural networks in hydrology. I: Preliminary

concepts, J. Hydrologic Eng., 52, 115–123

4. ASCE Task Committee on Application of Artificial Neural Networks in

Hydrology. 2000b. Artificial neural networks in hydrology. II: Hydrologic

applications, J. Hydrologic Eng., 52, 124–137

5. Behera, S., and Panda, R. K. (2006) Evaluation of management alternatives

for an agricultural watershed in a sub-humid subtropical region using a

physical process based model, Agric. Ecosyst. Environ., 113(1–4), 62–72

6. Bhattacharya, Sarangi, A., and A. K. (2005) Comparison of artificial neural

network and regression models for sediment loss prediction from Banha

watershed in India, Agric. Water Manage., 78(3), 195–208.

7. Cigizoglu, H. K. and Kisi, O. (2006) Methods to improve the neural network

performance in suspended sediment estimation, Journal of Hydrology, 317(3–

4), 221–238.

75
8. Cigizoglu, H. K. and Alp, M., (2007) Suspended Sediment load simulation

by two artificial neural network methods using hydro meteorological data,

Environmental Modelling & Software, January 2007, 2-13, 22(1), 2–13.

9. Cigizoglu, H. K. and Partal, T., (2008) Estimation and forecasting of daily

suspended sediment data using wavelet-neural networks, Journal of

Hydrology, 358(3–4), 317–331

10. De Farias, Alves, F. M., Santos, C. A. G., and Koichi, S. (2010) An ANN-

based approach to modelling sediment yield: a case study in a semi-arid area of

Brazil, Proc., ICCE Symp.: Sediment Dynamics for a Changing Future,

International Commission on Continental Erosion, Rennes, France, 337, 316–

321.

11. Duan, Z., Song, X., and Liu, J. (2009) Application of SWAT for sediment

yield estimation in a mountainous agricultural basin, 17th International

Conference on Geoinformatics, IEEE, 1–5.

12. Kaur, R., Srinivasan, R., Mishra, K., Dutta, D., Prasad, D., and Bansal, G.

(2003) Assessment of a SWAT model for soil and water management in India,

Land Use Water Resour. Res., 3, 1–7.

13. Mutlu, E., Chaubey, I., Hexmoor, H., and Bajwa, S. G. (2008) Comparison

of artificial neural network models for hydrologic predictions at multiple

gauging stations in an agricultural watershed, Hydrol. Processes, 22(26), 50–

97.

14. Nagraj S. Patil and Priyanka S. Desai (2016) Runoff and Sediment Yield

Modeling for Malaprabha Sub-Basin using SWAT Hydrological Model,

International Journal of Agricultural Science, v.7(2), pp.200-208.

76
15. Nagy, H.M., Watanabe, K., Hirano, M., (2002) Prediction of sediment load

concentration in rivers using Artificial Neural Network Model, Journal of

Hydraulic Engineering (ASCE) 128 (6), 588-595

16. Oeurng, C., Sauvage, S., Miguel, J., and Perez. S., (2011) Assessment of

hydrology, sediment and particulate organic carbon yield in a large agricultural

catchment using the SWAT model, Journal of Hydrology, 401(3-4):145-153

17. Prabhanjan A., E. P. Rao, and T. I. Eldho (2015) Application of SWAT

Model and Geospatial Techniques for Sediment-Yield Modeling in Ungauged

Watersheds, Journal of Hydrologic Engineering, 20(6): C6014005

18. Raghuwanshi N. S., Singh, R. and Reddy, L. S. (2006) Runoff and Sediment

Yield Modeling Using Artificial Neural Networks: Upper Siwane River, India,

Journal of Hydrologic Engineering, Vol. 11, No. 1, January 1, 2006,

10.1061/(ASCE)1084- 0699(2006)11:1(71).

19. Rajaee, T., Mirbagheri, S. A., Zounemat-Kermani, M., and Nourani, V.

(2009) Daily suspended sediment concentration simulation using ANN and

neuro-fuzzy models, Science of the Total Environment, 407(17), 4916–4927.

20. Setegn, G. S., Dargahi, B., Srinivasan, R. (2010) Modeling of sediment yield

from Anjeni-gauged watershed, Ethiopia using SWAT model, Journal of the

American Water Resources Association, Vol.3, 514-526.

21. Sumi, T. And Hirose, T. (2009) Accumulation of sediment in reservoir, in

water storage, Transport and Distribution, Encyclopaedia of Life Support

systems, Vol. 1, pg 489. ISBN 978-1-84826-176-1

22. Walling and Fang (2003) Recent trends in the suspended sediment loads of

the world’s rivers, Global and Planetary Change, 39 (2003) 111 – 126

77
23. Wang, Y. M., Kerh, T., and Traore, S. (2009) Neural networks approaches

for modelling river suspended sediment concentration due to tropical storms,

Global NEST J., 11(4), 457–466

24. Yenigün, K., Mahmut, B., Reşit, G., and Mehmet, M. (2010) A

comparative study on prediction of sediment yields in the Euphrates Basin,

Int. J. Phys. Sci., 5(5), 518–534.

25. Yesuf, H.M., Assen, M., Alamirew, T., and Melesse, M. (2015) Modeling

of sediment yield in Maybar gauged watershed using SWAT, northeast

Ethiopia, Catena, 127, 191–205

26. Zhu, Lu X. X, and Zhou Y. (2007) Suspended sediment flux modeling with

artificial neural network: An example of the Longchuanjiang River in the

Upper Yangtze Catchment, China, Geomorphology 84 (2007) 111–125

27. https://www.sciencedirect.com/science/article/pii/S1383730396800093

28. https://www.water.wa.gov.au/water-topics/waterways/threats-to-our-

waterways/erosion-and-sedimentation

29. http://cwc.gov.in/sites/default/files/hydrological-network-details-of-cwc.pdf

30. https://reliefweb.int/sites/reliefweb.int/files/resources/Rev-0.pdf

31. https://www.degruyter.com/view/journals/secm/20/4/article-

p319.xml?language=en

78