You are on page 1of 33

TABLE OF CONTENTS

1. ABSTRACT ....................................................................... ……………………………….3


2.INTRODUCTION ............................................................................................................. 4
2.1 Literature Review ...................................................................................................... 4
2.2 Objective And Approach ........................................................................................... 7
3.DATASET DETAILS ....................................................................................................... 8
3.1 Soil Moisture ............................................................................................................. 8
3.2 Temperature And Precipitation ................................................................................ 8
3.3 Study Area ................................................................................................................ 9
4.THEORY ......................................................................................................................... 9
4.1 Climate Elasticity .................................................................................................... 10
4.2 Sensitivity Analysis ................................................................................................. 10
4.3 Water Budget Equation .......................................................................................... 11
4.3.1 Evapotranspiration ....................................................................................... 11
4.3.2 Stream flow Discharge ................................................................................. 11
4.3.3 Groundwater Loss ....................................................................................... 12
4.4 Least Square Support Vector Machine .................................................................. 12
4.4.1 RBF Kernel .................................................................................................. 13
4.4.2 Model Calibration ......................................................................................... 14
4.4.3 ‘Leave-one-out’ Cross Validation ................................................................. 14

4.5 Performance Evaluation ............................................................................................ 14

5.METHODOLOGY .......................................................................................................... 15
5.1 Proposed Model ...................................................................................................... 15
5.1.1 Model 1 ......................................................................................................... 15
5.1.2 Model 2 ......................................................................................................... 15
5.1.3 Model 3 ......................................................................................................... 16
5.1.4 Model 4 ......................................................................................................... 16
5.2 Climate Elasticity ..................................................................................................... 16
5.3 Two Way Sensitivity Analysis .................................................................................. 17

1|Page
6.RESULTS ..................................................................................................................... 17
6.1 Proposed Model………………………………………………………………………......17
6.1.1 Model 1 ......................................................................................................... 17
6.1.2 Model 2 ......................................................................................................... 21
6.1.3 Model 3 ......................................................................................................... 23
6.1.4 Model 4 ......................................................................................................... 25
6.2 Climate Elasticity ..................................................................................................... 27
6.2.1 Temperature Elasticity .................................................................................. 28
6.2.2 Precipitation Elasticity ................................................................................... 30
6.3 Sensitivity Analysis.................................................................................................. 30
7. REFERENCE ........................................................................................................................ 31

2|Page
1. ABSTRACT

Accurate estimation of soil moisture is very important for agriculture based economies like India.

The main problem associated with estimating soil moisture is difficulty in obtaining the variables

such as evapotranspiration and groundwater loss, which are the basis of a water budget equation.

This study proposes four climate elasticity based models for estimating soil moisture using LSSVM

for non-parametric regression. The soil moisture estimation is entirely based upon precipitation and

temperature data. Using these models precipitation and climate elasticity have been obtained at 81

grid locations. The elasticity values obtained are a function of the climate variable which remains

constant for consecutive months. This gives more accurate estimation of change in soil moisture.

Using the models two way sensitivity analysis has also been done for soil moisture with respect to

temperature and rainfall.

3|Page
2. INTRODUCTION

Estimation can be of two types – parametric and non-parametric estimation. Most of the research

done in the field of soil moisture estimation is based upon the parametric approach. Although global

datasets of soil moisture are now available the error present in them is not negligible, With arrival

of different mathematical tools such a neural network, SVM’s, LSSVM etc.non-parametric

estimation of soil moisture is a potential research area, Estimation of soil moisture is mainly done

using in-situ data on a local scale without relating it with the climatic variables. In this work, LSSVM

has been used for estimation of soil moisture using climatic variable such as temperature and

rainfall as inputs. Also, attempt has been made to understand the individual impact of temperature

and precipitation on soil moisture. Climate elasticity has been approached differently in this work

and elasticity indices have been determined as function of some climatic variable.

2.1 Literature Review

Accurate estimation and prediction of soil moisture has been of interest to researchers since a long

time which still has not been achieved. This is mainly due to the fact that the standard estimation

protocol, i.e., the gravimetric method, which gives the most accurate result, is essentially a point

measurement and hence cannot be used to obtain gridded data.

The work of Rawls et al. (1982) is one of the most cited work in the field of estimation of soil

moisture. Their work resulted in assembling of extensive soil-water data and hydraulic conductivity

datasets which was later used to derive different parameters such as Brooks and Coery water

retention parameters, total porosity, soil water retention at different pressure and most importantly

relationships for predicting water retention volumes for certain parameters.

Margulis et al. (2001) has estimated the soil moisture over large areas by merging information

obtained from satellites, ground-based stations and regional hydrological model. The paper uses

Kalman Filtering, a data assimilation approach, for characterizing soil moisture and related land –

atmosphere fluxes. Similarly, Carson et al. (1990) has estimated the surface and root zone soil

4|Page
water content, surface energy fluxes and vegetation fraction using only satellite or remotely sensed

data.

Huang et al. (1996) is the basis of leaky bucket model used to calculate soil moisture data over a

given area using the water balance equation in soil. While modelling it uses the instantaneous soil

moisture data at the end of the month as its input while the predicted data is compared with the

observed monthly mean. The verification of runoff is not independent as calibration and verification

periods are the same. This paper has used both daily and monthly precipitation data for analysis

and observed that the results were quite similar. The Leaky Bucket Model is insensitive or not

overly sensitive to the details of when precipitation was falling.

Yoo et al. (2006) models the temporal behaviour of soil moisture using Poisson process model for

rainfall forcing to determine the mean, covariance and spectral density of soil moisture. The paper

also determines the sensitivity of soil moisture with respect to model parameters and observes that

soil moisture is mainly dependent upon arrival rate of rainfall rather than the amount of rainfall over

an area. However, the model is found to be insensitive to changes is model parameter as much as

20%. A better correlation is observed for longer training period.

Lakshmi et al. (2000) determined that along with soil moisture, surface temperature is a key factor

in determining the land surface heat and water balance. Upon the beginning of dry-down period of

soil, after a precipitation event there is a gradual decrease in the soil moisture and a sharp upward

trend in temperature.

Mils (1994) focuses on estimating soil moisture using the water budget equation in a 5km by 5km

grid of Ireland and has used a 30 year training dataset. Evapotranspiration has been calculated by

measuring the potential evapotranspiration (𝐸𝑝 ). Various methods of calculating 𝐸𝑝 are used whose

results are generally overestimates of evapotranspiration and are adjusted. Similar work for

estimating the soil moisture deficit and potential evaporation has been done by Burke and Alley

(1984) and Verstraeten et al. (2008). Actual evapotranspiration is calculated using 𝐸𝑝 and 𝑃 values.

5|Page
The 𝐸 values obtained in this paper are based upon temperature and precipitation values using

polynomial regression.

Sankarasubramanian and Vogel (2002) compared the performance of various models for

calculating the climate elasticity of soil moisture storage capacity of a basin. Some of the models

used are: (𝑖) conceptual deterministic watershed model, Schaake (1990), Nash and Gleick (1991)

and Jeton et al. (1996) have performed this type of study, (𝑖𝑖) multivariate regional hydrologic

models for different basins in a region. Similar work done has been done by Vogel et al., (1999)

(𝑖𝑖𝑖) empirical estimation of changes in stream flow due to changes in climate .Risbey and

Entekhabi, (1996).

Sankarasubramanian et al. (2001) determines the precipitation elasticity of stream flow using both

parametric and non-parametric approaches. Model Based estimates of 𝜀𝑝 are found to be highly

sensitive to error in model structure and calibration. Non parametric estimators are found to have

low bias and is as robust as or more robust as compared to model based approaches. This paper

highlight an example where different precipitation sensitivity of stream flow are calculated by two

different papers for a given site both of which have been verified by subsequent papers. This shows

that values of elasticity (𝜀𝑝 ) are also dependent upon the underlying assumptions of each model.

This paper defines climate elasticity of stream flow is defined by the proportional change in stream

flow Q by proportional change in a climatic variable such as precipitation P. Dooge (1992) has

calculated the sensitivity of surface runoff to precipitation using relation between relative changes

in precipitation, potential evapotranspiration and surface runoff.

Breierova and Choudhari (1996) focuses on parameter sensitivity and determining how sensitive a

model is to changes in its parameters. Sensitivity analysis has been used here to determine the

accuracy needed in a parameter to make a model sufficiently useful and valid.

Suykenss et al., (2002) has been used for formulation of Least Squares Support Vector Machines,

World Scientific Publishing C. Pvt. Ltd.

6|Page
2.2 Objective and Approach

The work done under this project is mainly concerned with prediction of soil moisture using

temperature and rainfall as input. It is assumed that evapotranspiration, surface runoff and

groundwater loss for a given area are function of temperature and precipitation.

The aim is to determine a relation between soil moisture, temperature and precipitation such that

we achieve a high correlation coefficient so that the trends can be predicted with more accuracy.

In this project, 4 models have been developed using LSSVM, which is a non-parametric learning

technique that analyses data and recognize patterns in data set which is used for regression. Using

trial and error method, the number of previous months’ data which affect soil moisture is determined.

Leave-one-out cross validation technique has been used while estimating the hyper parameters of

LSSVM model. It is a trade-off between the efficiency of the model and runtime.

Another objective of this project is to determine the precipitation and temperature elasticity of soil

moisture. However, instead of assuming that all the climatic variables, except the one whose effect

is under investigation, remain constant elasticity values determined are a function of these climatic

variables. For example, in this work the precipitation elasticity is a function of temperature and vice

versa. Non-parametric approach for estimating elasticity at the mean values of individual elasticity

at all data points. It has been observed that non parametric approach is more robust and accurate

as compared to parametric approach while estimating elasticity index. Model outputs are used for

estimation of elasticity indices in this project.

Sensitivity analysis has also been done for soil moisture with respect to temperature and

precipitation. It is a study of how the uncertainty or change in on or more climatic variables affect

the output of the model. This concept is mainly used in economics but is quite useful in the context

of this work also. It also allows us to know which variable has a higher impact on the soil moisture

and also the impact of simultaneous changes in both of them.

7|Page
For presenting the results of 2 way analysis a matrix is formed whose cells contain relative change

in soil moisture as its values and the axes represent change in temperature and precipitation

respectively.

The grid size for this study is 9 X 9 and models have been developed for each of the 81 grid points.

The training and testing period are varied to obtain the best possible model. For each model,

performance at all grid points have been calculated in terms of RMSE, MAE and correlation

coefficient.

3. DATASET DETAILS

All the datasets have been downloaded from Earth System Research Laboratory, Physical

Sciences Division, National Oceanic and Atmospheric Administration, US Department of

Commerce.

Website: http://www.esrl.noaa.gov/psd/data/gridded/

The details about study area and study period have been mentioned below:

3.1 Soil Moisture

• CPC Soil Moisture (V2)

• Time Period : Jan, 1980 to Jan, 2010

• Training Period : Jan 1980 to Dec 1999

• Testing Period : Jan 2000 to Jan 2010

• Temporal Resolution : Monthly Mean

• Spatial Coverage: 22.75˚N to 26.75˚N

82.25˚E to 86.25˚E

• Spatial Resolution: 0.5 ˚ X 0.5 ˚

• Unit: Soil Moisture – mm

3.2 Precipitation and Temperature

• University of Delaware Air Temperature & Precipitation

8|Page
• Time Period : Jan, 1980 to Jan, 2010

• Training Period : Jan 1980 to Dec 1999

• Testing Period : Jan 2000 to Jan 2010

• Temporal Resolution : Monthly Mean

• Spatial Coverage: 22.75˚N to 26.75˚N

82.25˚E to 86.25˚E

• Spatial Resolution: 0.5 ˚ X 0.5 ˚

• Unit: Temperature - ˚C, Rainfall - cm

3.3 Study Area

The study area lies within India having the same spatial coverage as the data above i.e.: 22.75˚N

to 26.75˚N and 82.25˚E to 86.25˚E. This has been marked by ‘x’ and shown in the figure below

Figure 1. Study area marked by ‘x’

4. THEORY

4.1 Climate Elasticity

9|Page
Climate elasticity of soil moisture is defined by the proportional change in soil moisture 𝑆 by

proportional change in a climatic variable such as precipitation 𝑃 [2]. Thus precipitation elasticity of

soil moisture is defined as

𝑑𝑆/𝑆 𝑑𝑄 𝑃
𝜀𝑝 (𝑃, 𝑆) = = (1)
𝑑𝑃/𝑃 𝑑𝑃 𝑄

Where 𝑆 is the soil moisture, 𝑃 is precipitation and 𝜀𝑝 is precipitation elasticity.

For simplicity climate elasticity can also be defined at mean value of a climate variable

𝑑𝑆 𝜇𝑃
𝜀𝑝 (𝜇𝑃 , 𝜇𝑆 ) = | (2)
𝑑𝑃 𝑃=𝜇𝑃 𝜇𝑆

Where 𝜇𝑆 is the mean soil moisture and 𝜇𝑃 is the mean precipitation

Another definition of climate elasticity of runoff with respect to precipitation is given by

∆𝑅 ∆𝑃 ∆𝑃𝐸
= 𝜓 − (𝜓 − 1) (3)
𝑅 𝑃 𝑃𝐸

Where 𝑅 is surface runoff and 𝑃𝐸 is potential evapotranspiration and 𝜓 is elasticity.

4.2 Sensitivity Analysis

Sensitivity analysis is used to determine how “sensitive” a model is to changes in the input variables

or the parameters of the model and to changes in model structure. Parameter sensitivity is usually

determined to see how a change in the value of a certain model parameter affects the behavior of

the model. It is an important tool in model building as well as evaluation.

Sensitivity analysis can be one-way or multi –way depending upon the kind of model. In one way

sensitivity analysis only one parameter is changed at one time and the variation in model output is

observed. This is same as estimating the climate elasticity for any parameter.

In multi-way sensitivity analysis more than one parameters of the model are simultaneously

changed and the change in output is observed. It gives us an idea about the underlying

relationship between model and parameters.

4.3 Water Budget Equation

10 | P a g e
This paper is the basis of leaky bucket model used to calculate soil moisture data over a given area

using the water balance equation in soil given by :

𝑑𝑆
= 𝑃(𝑡) − 𝐸(𝑡) − 𝐺(𝑡) − 𝑅(𝑡) (4)
𝑑𝑡

Where 𝑊 is the soil water content, 𝑃 is precipitation, 𝐸 is evapotranspiration, 𝐺 is ground water loss

through deep percolation and 𝑅 is net stream flow divergence.

4.3.1 Evapotranspiration (𝐸)

It is calculated using the relation


𝑊
𝐸 = 𝐸𝑝 𝑊 (5)
𝑚𝑎𝑥

Where 𝐸𝑝 represents Potential evapotranspiration which is the maximum and 𝑊𝑚𝑎𝑥 is the measure

of soil capacity to hold water in mm.

𝐸𝑝 mainly depends upon the net radiation heating of the surface. To meet this demand, vegetation

uses water available through precipitation and/or water stored in the soil (𝑊) depending upon the

amount of precipitation occurring over the area. When 𝑃 > 𝐸𝑝 there is sufficient water and actual

evapotranspiration (𝐸) will be equal to 𝐸𝑝 and the extra water either contributes to increased soil

storage or to surplus. If the opposite occurs i.e. 𝑃 < 𝐸𝑝 , there is insufficient water to meet 𝐸𝑝

requirements. To meet this demand water stored in the soil will be extracted by the root system

4.3.2 Stream flow Divergence (𝑅)

The stream flow divergence 𝑅(𝑡) consists of a surface runoff component 𝑆(𝑡) and a subsurface

(base flow) runoff component 𝐵(𝑡) given by:

𝑊(𝑡) 𝑚
𝑆(𝑡) = 𝑃(𝑡) [𝑊 ] (5)
𝑚𝑎𝑥


𝐵(𝑡) = 𝑊(𝑡) (6)
1+𝜇

Where m is a parameter with values greater than 1, ∝ is the inverse of response time of the baseflow

and 𝜇 is a dimensionless parameter that determines the portion of the subsurface flow that becomes

base flow in the channels draining out from the area.

11 | P a g e
4.3.3 Groundwater Loss (𝐺)

It is obtained by the relation

𝜇∝
𝐺(𝑡) = 𝑊(𝑡) (7)
1+𝜇

4.4 LSSVM Formulation

Consider a model in the primal weight space of the following form:

𝑦(𝑥) = 𝜔𝑇 ∅(𝑥) + 𝑏 (8)

Where x ∈ Rn and y ∈ Rn and ∅( . ): Rn → Rnh is the mapping to the high dimensional and

potentially infinite dimensional feature space. For a given training set { 𝑥𝑘 , 𝑦𝑘 }𝑁


𝑘=1 we can

formulate the following optimization problem in the primal weight space


𝑁
1 1
min 𝐽𝑃 (𝜔, 𝑒) = 𝜔𝑇 𝜔 + 𝛾 ∑ 𝑒𝑘2 (9)
2 2
𝑘=1

Such that,

𝑦(𝑥) = 𝜔𝑇 ∅(𝑥) + 𝑏 + 𝑒𝑘 𝑘 = 1, 2, 3 … . 𝑁 (10)

When 𝑤 becomes infinite dimensional, this primal problem cannot be solved. Hence, lagrangian

of this problem is formed and dual problem is derived,


N

𝐿(𝜔, 𝑏, 𝑒 ; 𝛼) = JP (𝜔, e) − ∑ αk { 𝜔T ∅(x) + b + ek − yk } (11)


k=1

where αk are Lagrangian Multipliers.

The conditions for optimality are given by,

N
𝜕𝐿
= 0 → 𝜔 = ∑ αk ∅(x) (12)
𝜕𝜔
k=1

N
𝜕𝐿
= 0 → ∑ αk = 0 (13)
𝜕𝑏
k=1

𝜕𝐿
= 0 → αk = γ𝑒𝑘 (14)
𝜕𝑒𝑘

12 | P a g e
𝜕𝐿
= 0 → 𝜔T ∅(x) + b + ek − yk = 0 (15)
𝜕αk

Where,

𝑦 = [y1 , y2 … yN ]

1𝑣 = [1, 1 … 1]

𝛼 = [α1 , α2 … αN ] (16)

After elimination of variables 𝜔 and e, we obtain the values of 𝑏 𝑎𝑛𝑑 𝛼 by solving the linear system.

‘𝜔’ is replaced by kernel function as follows:

𝐾(𝑥, 𝑥𝑖 ) = ∅(xi )∅(x)T (17)

This is called ‘kernel trick’. It is a way to map the observations to an inner product space without

actually computing the mapping and as per expectations the observation will have meaningful linear

structure in that inner product space.

Thus, the resulting LSSVM model can be expressed as,


N

𝑦(𝑥) = ∑ αk K(𝑥, 𝑥𝑘 ) + b (18)


k=1

where, K(𝑥, 𝑥𝑘 ) is the Kernel function.

4.4.1 RBF Kernel

In comparison with other kernel functions such as polynomial kernels, linear kernel function etc. ,

Radial Basis Function (RBF) kernel is more compact and able to shorten the computational training

process and improve the generalization performance of LS – SVR, a feature of great importance in

designing a model (8). The non-linear radial basis function kernel is defines as,

1
𝐾(𝑥, 𝑥𝑖 ) = exp (− ‖𝑥 − 𝑥𝑖 ‖) (19)
𝜎2

where 𝜎 is the kernel function parameter of RBF kernel. The symbol ‖ . ‖ is the norm of the vector

∙ and thus, ‖𝑥 − 𝑥𝑖 ‖2 is the Euclidian distance between 𝑥 and 𝑥𝑖 . In context of the stream flow

prediction 𝑥𝑖 is the new vector of previous stream flow, based upon which prediction 𝑦𝑖 is made.

13 | P a g e
4.4.2 Model Calibration

The regularization parameter 𝛾 represents a trade-off between the fitting error minimization and the

smoothness of the curve. For the given problem 𝜎, 𝛾 are determined using trial and error methods

to achieve maximum performance with LS-SVR models. Thus both the parameters are calibrated

during the training period. In this project the validation technique used with the training dataset is

Leave one out method. The length of training dataset is varied to obtain the best fit between

predicted and observed data.

4.4.3 ‘Leave-one-out’ Cross Validation

Cross validation is a model validation technique which is used to assess the performance of a model

for independent data set. It is used when the model is mainly used for prediction and the accuracy

of a model is to be determined. As the name suggests, leave one out cross validation technique

uses a single observation from the training dataset as the validation data and the remaining

observations as the training data. This process is repeated till each observation in t data set has

been used as validation data.

4.5 Performance Evaluation

The performance of the model on both the training data and the testing data are measured using

Root Mean Square Error (RMSE), Mean Absolute Error and Correlation Coefficient (R), which are

widely used for evaluating results of time series forecasting.

𝑁
1
𝑅𝑀𝑆𝐸 = √ ∑(𝑦𝑖 − 𝑦𝑜 )2 (20)
𝑁
1

𝑀𝐴𝐸 = ∑𝑁
1 |𝑦𝑖 − 𝑦𝑜 | (21)

1
𝑛 (𝑦𝑖 − 𝑦̅)(𝑦
𝑖 𝑜 − ̅̅̅)
𝑦𝑜
𝑅= (22)
√ 1 ∑𝑁 2 √ 1 ∑𝑁 𝑦𝑜 2
𝑁 1 (𝑦𝑖 − 𝑦̅𝑖 ) 𝑁 1 (𝑦𝑜 − ̅̅̅)

5. METHODOLOGY

14 | P a g e
In this project non-parametric estimation of monthly soil moisture has been done using Least

Square Support Vector Machine with monthly mean temperature and rainfall values as input.

5.1 Proposed Model

Total 4 models have been developed during this work out of which 1 model is used for prediction

of soil moisture and the rest are used for predicting the percentage change in soil moisture. ‘LS-

SVMlab’, a Matlab toolbox under Windows has for nonparametric regression estimation based on

LSSVM for this project.

5.1.1 Model 1

This model is used for predicting relative change in soil moisture and takes previous two months’

temperature and precipitation data as input.

∆𝑆𝑖+1
= 𝑓(𝑇𝑖+1 , 𝑅𝑖+1 , 𝑇𝑖 , 𝑅𝑖 ) (23)
𝑆𝑖

𝑆𝑖+1 = 𝑆𝑖 + ∆𝑆𝑖+1 (24)

5.1.2 Model 2

This model is used for predicting relative change in soil using the signal hidden in temperature of

soil moisture 𝑔(𝑇), and signal hidden in precipitation for soil moisture ℎ(𝑃). The functions 𝑔 and ℎ

are estimated using LSSVR approach by individually using precipitation and temperature to predict

soil moisture. The individual dependence of soil moisture on temperature and precipitation is

determined by using first half of training dataset. In the next half, relative changes in these functional

values based upon the previous step value are used as input to model relative change in soil

moisture.

∆𝑆𝑖+1 ∆𝑔(𝑇𝑖+1 ) ∆ℎ(𝑃𝑖+1 )


= 𝑓( , ) (25)
𝑆𝑖 𝑔(𝑇𝑖 ) ℎ(𝑃𝑖 )

15 | P a g e
𝑆𝑖+1 = 𝑆𝑖 + ∆𝑆𝑖+1 (26)

5.1.3 Model 3

This model is used for predicting relative change in soil moisture based upon the relative change in

temperature as well as precipitation data. The resultant model is used to perform two – way

sensitivity analysis of soil moisture with respect to temperature and rainfall.

∆𝑆𝑖+1 ∆𝑇𝑖+1 ∆𝑃𝑖+1


= 𝑓( , ) (27)
𝑆𝑖 𝑇𝑖 𝑃𝑖

𝑆𝑖+1 = 𝑆𝑖 + ∆𝑆𝑖+1 (28)

5.1.4 Model 4

This model is used for predicting soil moisture by taking previous ‘n’ months’ temperature and

precipitation data as input.

𝑆𝑖+1 = 𝑓(𝑇𝑖+1 , 𝑅𝑖+1 , . . . . 𝑇𝑖−𝑛 , 𝑅𝑖−𝑛 ) (29)

5.2 Climate Elasticity

In this project non parametric estimation of climate elasticity of soil moisture is done with respect

to (𝑖) precipitation and (𝑖𝑖) temperature. In non-parametric estimation elasticity is taken at the

mean values of individual elasticity at all the data points:

𝑛
1 𝑑𝑆 𝜇𝑃
̅𝜀̅𝑝̅̅ (𝜇𝑃 , 𝜇𝑆 ) = ∑ | (30)
𝑛 𝑑𝑃 𝑃=𝜇𝑃 𝜇𝑆
𝑖=1

Where 𝜀𝑝 is precipitation elasticity of soil moisture

𝑛
1 𝑑𝑆 𝜇𝑇
𝜀𝑇 (𝜇𝑇 , 𝜇𝑆 ) = ∑ |
̅̅̅ (31)
𝑛 𝑑𝑇 𝑃=𝜇𝑃 𝜇𝑆
𝑖=1

Where 𝜀𝑇 is temperature elasticity of soil moisture

5.3 Two-Way Sensitivity Analysis

16 | P a g e
For performing a two way analysis both temperature and precipitation was simultaneously varied

and the relative change in soil moisture value was observed using Model 1. The results are stored

in a matrix with the axes representing relative change in temperature, relative change in

precipitation and the intersection containing relative change in soil moisture.

6. RESULTS AND DISCUSSION

6.1 Proposed Model

A total of 240 observations of monthly temperature and soil moisture were used to train the LSSVM

model. Validation of the model was carried out on 120 observation. The results were compared

based upon the root mean square error for each model. Individual models are developed for all the

81 grid points and the performance is compared using RMSE, MAE and correlation coefficient

All the graphs and table below have been obtained for the grid point 86.25°E and 23.75°N for

a period of 1980 – 2000 unless mentioned otherwise.

6.1.1 Model 1

This model is used to obtain relative change in soil moisture. By trial and error method it is observed

that the RMSE for n=4 is the least and has the best correlation value. This is in contrast with the

water budget equation where the change in soil moisture is modelled using only previous two

monthly values of rainfall and temperature. Given below are the results for 81 grid points in the

study area (n=4)

Table 1. Performance evaluation at 81 grid points in terms of Mean Absolute Error (MAE), Root

Means Square Error (RMSE) and Correlation Coefficient (R) of Model 1

0.05956 0.04676 0.05089 0.05726 0.04686 0.04659 0.03644 0.0314 0.03352

0.07498 0.05949 0.04961 0.05173 0.04008 0.03575 0.03272 0.03135 0.05115


MAE
0.07907 0.06346 0.04986 0.0479 0.03698 0.03291 0.03019 0.03112 0.05009

0.07561 0.05787 0.04161 0.03767 0.03753 0.04221 0.03594 0.02884 0.03182

17 | P a g e
0.07073 0.0473 0.03627 0.03077 0.03765 0.04773 0.03614 0.02879 0.04104

0.06933 0.03876 0.03311 0.03036 0.03569 0.03929 0.03245 0.02885 0.02798

0.06971 0.04072 0.03645 0.03306 0.03274 0.03498 0.03136 0.0321 0.03378

0.07066 0.04788 0.03526 0.03165 0.02904 0.03157 0.03138 0.03382 0.04184

0.07007 0.0602 0.05345 0.04416 0.03794 0.03312 0.03674 0.04553 0.05427

0.08304 0.06697 0.07197 0.07994 0.06519 0.06349 0.0512 0.04332 0.04512

0.1069 0.08228 0.06922 0.06811 0.05541 0.0491 0.04591 0.04378 0.04261

0.11716 0.09326 0.0694 0.06737 0.05444 0.04915 0.04216 0.04205 0.03492

0.11042 0.08123 0.05757 0.05263 0.05869 0.06639 0.0514 0.03759 0.04383

RMSE 0.10229 0.06795 0.04994 0.04175 0.05511 0.07458 0.05253 0.03705 0.04119

0.09329 0.05516 0.04393 0.04062 0.04911 0.05553 0.04273 0.03626 0.03626

0.09539 0.0568 0.05047 0.04524 0.04587 0.04766 0.04235 0.04061 0.04502

0.10006 0.06495 0.05084 0.04339 0.04065 0.04527 0.04189 0.04703 0.06256

0.10122 0.08479 0.07803 0.06406 0.05257 0.04574 0.05402 0.06787 0.08369

0.94507 0.9635 0.95134 0.93858 0.95959 0.96058 0.97162 0.97702 0.97653

0.92059 0.94654 0.94799 0.94915 0.9652 0.97309 0.97454 0.97672 0.97915

0.90167 0.92594 0.94828 0.94522 0.96266 0.97159 0.97896 0.97896 0.97834

0.90512 0.94159 0.96578 0.9689 0.96215 0.95049 0.96873 0.98416 0.97969

R 0.92079 0.95947 0.97517 0.98111 0.9698 0.94322 0.96877 0.98519 0.98214

0.92879 0.97235 0.98155 0.98345 0.97589 0.97057 0.97966 0.98452 0.98592

0.91779 0.97255 0.97688 0.97996 0.97922 0.9777 0.97995 0.98079 0.97801

0.89223 0.96282 0.97738 0.98251 0.98378 0.97965 0.98021 0.97465 0.95991

0.88382 0.93202 0.94601 0.96342 0.97287 0.97911 0.97186 0.95176 0.9277

Table 2. Performance evaluation for different number of inputs (n) of Model 1 at point (3, 9)

18 | P a g e
n=2 n=3 n=4 n=5

RMSE 0.0480 0.0429 0.0336 0.0389

MAE 0.0787 0.0727 0.0553 0.0584

R 0.9556 0.9663 0.9789 0.9713

Figure 1. Observed Change versus Predicted change in soil moisture for Jan 2000 to Jan 2012

for two month precipitation and temperature data as inputs

19 | P a g e
Figure 2. Observed Change versus Predicted change in soil moisture for Jan 2000 to Jan 2012

for three month precipitation and temperature data as inputs

Figure 3. Observed Change versus Predicted change in soil moisture for Jan 2000 to Jan 2012

for four month precipitation and temperature data as inputs

Figure 4. Observed Change versus Predicted change in soil moisture for Jan 2000 to Jan 2012

for five month precipitation and temperature data as inputs

20 | P a g e
6.1.2 Model 2

This model is uses LSSVM in two stages i.e. output of one LSSVM is used as the input for other

LSSVM. The main concept behind modelling the temperature and rainfall cause before using them

for developing final model is that both of these variable have some individual non- linear relationship

with soil moisture which is evident from the scatter plots. Even though unsatisfactory, modelling

them individually with soil moisture helps in obtaining that non-linear relationship to some extent.

Table 3. Performance evaluation at 81 grid points in terms of Mean Absolute Error (MAE), Root

Means Square Error (RMSE) and Correlation Coefficient (R) of Model 2

0.21596 0.19667 0.16115 0.14927 0.15773 0.2026 0.23417 0.24581 0.19212

0.28698 0.19268 0.19226 0.1607 0.16455 0.18627 0.17383 0.21395 0.2092

0.3279 0.26068 0.19441 0.15557 0.17362 0.14677 0.15394 0.21454 0.16958

0.28951 0.25161 0.22204 0.18729 0.15413 0.23448 0.18221 0.25417 0.26158

MAE 0.29714 0.30108 0.2428 0.16053 0.1621 0.27975 0.13657 0.17351 0.22341

0.21894 0.28441 0.29675 0.23198 0.26467 0.16399 0.19044 0.1561 0.2089

0.2185 0.27426 0.28733 0.27396 0.22631 0.17552 0.20143 0.18799 0.24953

0.26516 0.2022 0.26529 0.1784 0.17381 0.18545 0.185 0.22322 0.25997

0.20792 0.17 0.25165 0.25482 0.22237 0.20996 0.2623 0.19874 0.2646

0.13248 0.12566 0.10575 0.10057 0.10797 0.12884 0.13054 0.14888 0.11491

0.15966 0.12737 0.12311 0.1053 0.11025 0.11641 0.10183 0.14414 0.1377

0.18394 0.14996 0.13016 0.10758 0.11009 0.08721 0.08596 0.12184 0.11292

RMSE 0.16787 0.14661 0.13601 0.11661 0.1047 0.12958 0.10032 0.14056 0.1499

0.16267 0.16894 0.1472 0.09761 0.09748 0.14624 0.07833 0.0996 0.12814

0.12191 0.16142 0.17087 0.14237 0.14584 0.10159 0.11604 0.09162 0.12361

0.12943 0.15813 0.16309 0.15451 0.13749 0.10039 0.12466 0.109 0.14415

21 | P a g e
0.14186 0.11976 0.14991 0.10805 0.11055 0.11472 0.10973 0.13148 0.15142

0.11582 0.10194 0.14272 0.14913 0.13615 0.12587 0.15661 0.12094 0.15307

0.59127 0.53946 0.61768 0.6376 0.58413 0.35239 0.48537 0.40399 0.72834

0.42002 0.56299 0.32818 0.50164 0.53967 0.49235 0.73091 0.50808 0.62996

0.284 0.44957 0.47137 0.52837 0.46423 0.78767 0.82521 0.67092 0.77245

0.41941 0.48666 0.46313 0.5664 0.69558 0.45171 0.74808 0.5016 0.5474

R 0.45081 0.26797 0.41329 0.77242 0.77832 0.31219 0.87803 0.79867 0.69281

0.74171 0.38052 0.30117 0.4398 0.3667 0.77093 0.68919 0.8343 0.71385

0.68221 0.5034 0.34691 0.23002 0.45873 0.769 0.63732 0.78192 0.56407

0.50011 0.73625 0.51961 0.73606 0.7 0.70476 0.75895 0.65142 0.54072

0.57751 0.80265 0.50582 0.34661 0.54916 0.66271 0.42427 0.76145 0.53912

Table 4. Performance evaluation for different number of inputs (n) of Model 1 at point (3, 9)

RMSE MAE R Duration

0.1174 0.1709 0.8027 2000 – 2010

22 | P a g e
Figure 5. Observed Change versus Predicted change in soil moisture for Jan 2000 to Jan 2012

using change in precipitation and temperature data as inputs

6.1.3 Model 3

This model simply uses relative change for modelling the relative change in soil moisture. The

performance of this model is better compared to that of model 2 which shows that the assumption

of a hidden signal within temperature and soil moisture is not very accurate.

Figure 6. Observed Change versus Predicted change in soil moisture for Jan 2000 to Jan 2012

with relative change in precipitation and temperature data as inputs

Table 5. Performance evaluation at 81 grid points in terms of Mean Absolute Error (MAE), Root

Means Square Error (RMSE) and Correlation Coefficient (R) of Model 3

0.46435 0.26102 0.26718 0.27481 0.27872 0.26024 0.23383 0.21708 0.22101

0.30956 0.23756 0.25311 0.27721 0.27336 0.25891 0.20033 0.1916 0.17366


MAE
0.34239 0.24938 0.24632 0.25893 0.25733 0.21582 0.1839 0.16617 0.15898

0.27495 0.25705 0.26833 0.26741 0.24822 0.22404 0.21772 0.18321 0.16686

23 | P a g e
0.27641 0.27286 0.26061 0.26231 0.25827 0.29932 0.24336 0.19261 0.17033

0.32796 0.27311 0.28514 0.2728 0.28317 0.2631 0.20726 0.21002 0.18845

0.34966 0.27882 0.32334 0.29948 0.30549 0.24064 0.22589 0.236 0.21628

0.3605 0.29149 0.24958 0.22466 0.25219 0.2478 0.23716 0.25121 0.27762

0.37058 0.30666 0.30341 0.23428 0.25251 0.25894 0.22675 0.28466 0.34369

0.6059 0.34081 0.36574 0.3935 0.37513 0.36816 0.31096 0.27991 0.28562

0.41004 0.31688 0.34844 0.40188 0.3847 0.37137 0.26174 0.24004 0.22486

0.47412 0.33402 0.32544 0.36015 0.35418 0.29958 0.24683 0.20664 0.19739

0.36219 0.33157 0.33592 0.34579 0.31716 0.29831 0.29223 0.23306 0.22037

RMSE 0.36262 0.34106 0.33466 0.31911 0.31825 0.37172 0.31725 0.23986 0.2291

0.4134 0.35418 0.36143 0.35449 0.36846 0.33748 0.26995 0.26936 0.2443

0.44481 0.37052 0.43741 0.40534 0.39176 0.30722 0.29613 0.31299 0.27651

0.45958 0.37168 0.33374 0.31049 0.32677 0.30962 0.31747 0.35336 0.35119

0.4702 0.4237 0.41512 0.31859 0.3381 0.33591 0.31628 0.37622 0.4681

0.64329 0.83485 0.76036 0.71812 0.7837 0.79349 0.86908 0.89719 0.89179

0.83489 0.86037 0.78889 0.74141 0.77379 0.80956 0.90557 0.9253 0.93307

0.76511 0.85515 0.83257 0.77837 0.79048 0.87831 0.9145 0.94345 0.94773


R
0.87143 0.86645 0.84361 0.83068 0.85286 0.88046 0.88429 0.9238 0.93817

0.87382 0.86638 0.86729 0.86095 0.86176 0.81899 0.86542 0.9245 0.93684

0.82477 0.86414 0.85633 0.84866 0.82499 0.84634 0.90207 0.90842 0.92817

24 | P a g e
0.78321 0.86219 0.81944 0.81205 0.81986 0.87413 0.89065 0.88996 0.91138

0.76259 0.87016 0.88571 0.89214 0.87022 0.87614 0.87909 0.86077 0.8736

0.74488 0.82396 0.82433 0.88605 0.86287 0.8674 0.88402 0.84559 0.78502

6.1.4 Model 4

This model is used for modelling the monthly soil moisture based upon previous ‘n’ months’

temperature and precipitation data similar to that in model 1. Similar to the previous result least

RMSE and best correlation is observed for n = 4. Soil moisture used in this model is normalized

with mean and standard deviation of the training period only.

Table 6. Performance evaluation for different number of inputs (n) of Model 1 at point (3, 9)

n=2 n=3 n=4 n=5

RMSE 0.1590 0.1003 0.0863 0.0865

MAE 0.1975 0.1306 0.1121 0.1131

R 0.9477 0.9777 0.9853 0.9833

25 | P a g e
Figure 7. Observed soil moisture (normalized) versus Predicted soil (normalized) for Jan 2000

to Jan 2012 for two month precipitation and temperature data as inputs

Figure 8. Observed soil moisture (normalized) versus Predicted soil (normalized) for Jan

2000 to Jan 2012 for three month precipitation and temperature data as inputs

26 | P a g e
Figure 9. Observed soil moisture (normalized) versus Predicted soil (normalized) for Jan

2000 to Jan 2012 for four month precipitation and temperature data as inputs

Figure 10. Observed soil moisture (normalized) versus Predicted soil (normalized) for Jan

2000 to Jan 2012 for five month precipitation and temperature data as inputs

6.2 Climate Elasticity

For estimating elasticity model 1 has been used with n=2 since we need to determine the impact of

change of a climate variable for two consecutive months. For determining precipitation elasticity it

is assumed that the temperature remains constant for two consecutive months while there is change

in the precipitation. For each temperature values elasticity is obtained. Elasticity values then plotted

against the corresponding temperature and hence, we obtain a temperature dependent elasticity

index for soil moisture. Similarly, temperature elasticity is obtained by assuming a constant

precipitation for two consecutive months and plotting the elasticity value against monthly

precipitation.

27 | P a g e
6.2.1Temperature Elasticity

Figure 11. Temperature elasticity of soil moisture as a function of precipitation which is assumed

to be constant for consecutive months

The graph show above represents the temperature elasticity of soil moisture as a function of rainfall.

Initially when the rainfall is low, the elasticity index is negative showing that in case of increase in

temperature there is reduction in soil moisture. This is as per expected lines because in case of low

precipitation the vegetation uses up the soil moisture and there are losses due to evaporation. At

zero rainfall for two consecutive month elasticity index is -0.43 which means that for every 10%

increase in temperature there is 4.3% reduction in soil moisture. Elasticity index is zero for 25 mm

rainfall, suggesting that irrespective of any temperature change there is no change in soil moisture

over the area. Beyond that the elasticity index is positive which shows that even if there is

temperature increase the amount of rainfall is sufficient to nullify the loss due to evaporation and at

the same time adds up to the existing soil moisture. For very high rainfall (>80cm) it can be observed

that the elasticity index tends to a constant value ~ 1.2. This happens because when the soil

28 | P a g e
moisture equals its field capacity, i.e. the soil is completely saturated any extra rainfall is treated as

surface runoff.

6.2.2 Precipitation Elasticity

Figure 12. Precipitation elasticity of soil moisture as a function of temperature which is assumed

to be constant for consecutive months

The graph represents the precipitation elasticity of soil moisture as a function of temperature.

Initially when the temperature is low the elasticity index is positive. In case the temperature is

constant at 10℃ for two consecutive month and the change in precipitation is 10%, the soil moisture

increases by 1%. This seems to be quite low which shows that sensitivity of soil moisture to change

in precipitation is low. At temperature around 25℃ the elasticity index is zero, meaning that for any

change in precipitation there is no change in soil moisture. For temperature greater than 25℃ the

elasticity index is negative showing that even if there is increase in precipitation there is a drop in

soil moisture. This probably happens because in case of high temperature the effect of evaporation

on soil moisture dominates over the effect of precipitation.

29 | P a g e
6.3 Sensitivity Analysis

Table 7. Two way sensitivity analysis of soil moisture with respect to temperature and precipitation

at point (3, 9)

Percent Change in Precipitation

-100 -75 -50 -25 0 25 50 75 100

-30 -23.9979 -21.0067 -25.125 -16.338 2.4411 22.01056 44.15622 30.02231 27.52635

-25 -24.7951 -20.5509 -24.3918 -16.6834 -2.44896 18.42975 42.3756 28.63339 29.81681

-20 -21.9879 -16.6597 -20.3963 -14.1 -5.86222 14.31135 38.00635 26.10221 31.77245

-15 -15.3486 -9.66892 -13.8919 -9.17071 -7.49747 10.02131 31.38798 22.79195 33.2247

-10 -5.41963 -0.69851 -6.38002 -3.24565 -7.65367 5.839653 23.25975 19.19604 34.03622
Percent change in temperature

-5 6.518972 8.582876 0.282794 1.932329 -7.10436 1.939356 14.62252 15.81547 34.11821

0 18.71324 16.39903 4.445989 4.794994 -6.76701 -1.57213 6.536239 13.04269 33.44512

5 29.35545 21.36409 5.18959 4.490432 -7.31142 -4.58366 -0.09039 11.08986 32.06375

10 37.02374 22.85629 2.605158 1.174534 -8.88433 -6.93335 -4.66149 9.978159 30.09294

15 40.99424 21.1349 -2.27502 -4.06756 -11.07 -8.41749 -6.96961 9.580467 27.71077

20 41.332 17.17411 -7.86138 -9.62783 -13.0895 -8.85371 -7.17659 9.691572 25.12943

25 38.75536 12.2989 -12.539 -13.93 -14.1327 -8.1638 -5.71871 10.09705 22.56261

30 34.34913 7.773313 -15.1396 -15.9073 -13.6654 -6.43049 -3.17254 10.62053 20.19327

The table shown above represents the two- way sensitivity analysis of soil moisture with respect to

temperature and precipitation. The axes represent the percent change in temperature and the

precipitation while the value at intersection represent the estimated soil moisture change. Model 1

has been used for obtaining this table. Upon comparing the results against the actual percent

change in soil moisture, it was observed that the results are good for cases where the change is

30 | P a g e
soil moisture is greater than 30%. This is explained by the graph obtained from Model 1. The

predicted and observed curve are close when in case of peaks but when the relative change in soil

moisture is low, there is considerable error. In order to improve upon this result we may use the

outputs of Model 4 may be used.

Upon comparing the overall performance of all the Models it is clear that the performance of Model

1 is the best. It is evident from the results that temperature and precipitation affect the soil moisture

for next four months. Poor performance of model 2 and 3 as compared to model 1 arise because

they assume that change in soil moisture is a function of only past two months while it is evident

from model 1 that best performance is observed for n=4. It was assumed that there is a soil moisture

signal hidden in precipitation and temperature values but the poor performance of model 2 as

compared to model 3 suggests otherwise. However as mentioned before the correlation is poor for

lower values of change in soil moisture in two way. Model 4 uses past n month temperature and

precipitation data to predict soil moisture and similar to model 1 it is observed that the best

correlation is observed for n=4.

7. REFERENCE

1. Vogel, R. M., & Sankarasubramanian, A. (2003). Validation of a watershed model without

calibration. Water resources research, 39(10).

2. Breierova, L., & Choudhari, M. (1996). An introduction to sensitivity analysis. Prepared for the

MIT System Dynamics in Education Project.

3. Frederick, K. D., & Major, D. C. (1997). Climate change and water resources. Climatic Change,

37(1), 7-23.

4. Carlson, T. N., Perry, E. M., & Schmugge, T. J. (1990). Remote estimation of soil moisture

availability and fractional vegetation cover for agricultural fields. Agricultural and Forest

Meteorology, 52(1), 45-69.

31 | P a g e
5. Margulis, S. A., McLaughlin, D., Entekhabi, D., & Dunne, S. (2002). Land data assimilation and

estimation of soil moisture using measurements from the Southern Great Plains 1997 Field

Experiment. Water resources research, 38(12), 35-1.

6. Sankarasubramanian, A., & Vogel, R. M. (2002). Annual hydroclimatology of the United States.

Water Resources Research, 38(6), 19-1.

7. Sankarasubramanian A., Vogel, R. M., & Limbrunner, F. J. (2001). Climate elasticity of

streamflow in the United States. Water Resources Research, 37(6), 1771-1781

8. Huang, J., van den Dool, H. M., & Georgarakos, K. P. (1996). Analysis of model-calculated soil

moisture over the United States (1931-1993) and applications to long-range temperature

forecasts. Journal of Climate, 9(6), 1350-1362.

9. Yoo, C., Kim, S. J., & Valdes, J. B. (2005). Sensitivity of soil moisture field evolution to rainfall

forcing. Hydrological processes, 19(9), 1855-1869.

10. Lakshmi, V., Zehrfuhs, D., & Jackson, T. (2000). Observations of land surface temperature and

its relationship to soil moisture during SGP99. In Geoscience and Remote Sensing Symposium,

2000. Proceedings. IGARSS 2000. IEEE 2000 International (Vol. 3, pp. 1256-1258). IEEE.

11. Mills, G. (2000). Modelling the water budget of Ireland—evapotranspiration and soil moisture.

Irish geography, 33(2), 99-116.

12. Burke, W. (1962). Calculated Potential Evapotranspiration and Soil Moisture Deficits for North

County Dublin (1950-1961). Irish Journal of Agricultural Research, 329-333.

13. Verstraeten, W. W., Veroustraete, F., & Feyen, J. (2008). Assessment of evapotranspiration

and soil moisture content across different scales of observation. Sensors, 8(1), 70-117.

14. Alley, W. M. (1984). On the treatment of evapotranspiration, soil moisture accounting, and

aquifer recharge

15. Dooge, J. C. (1992). Sensitivity of runoff to climate change: A Hortonian approach. BULLETIN-

AMERICAN METEOROLOGICAL SOCIETY, 73, 2013-2013.

32 | P a g e
16. Hakami, A., Odman, M. T., & Russell, A. G. (2003). High-order, direct sensitivity analysis of

multidimensional air quality models. Environmental science & technology, 37(11), 2442-2452.

17. Pelckmans, K., Suykens, J. A., Van Gestel, T., De Brabanter, J., Lukas, L., Hamers, B., &

Vandewalle, J. (2002). LS-SVMlab: a matlab/c toolbox for least squares support vector

machines. Tutorial. KULeuven-ESAT. Leuven, Belgium.

18. Schaake, J. C., & Waggoner, P. E. (1990). From climate to flow. Climate change and US water

resources., 177-206.

19. Nash, L. L., & Gleick, P. H. (1991). Sensitivity of streamflow in the Colorado basin to climatic

changes. Journal of hydrology, 125(3), 221-241.

20. Jeton, A. E., Dettinger, M. D., & Smith, J. L. (1996). Potential effects of climate change on

streamflow, eastern and western slopes of the Sierra Nevada, California and Nevada. US

Department of the Interior, US Geological Survey.

33 | P a g e

You might also like