You are on page 1of 64

CROP YIELD ESTIMATION OF WHEAT BY INTEGRATING

REMOTE SENSING, LAND AND MANAGEMENT FACTORS

A case study of Saharanpur District, Uttar Pradesh, India

A Project Report Submitted in Partial Fulfillment of The Requirements for The Award
of Post Graduate Diploma in Remote Sensing and Geographical Information System

By :
THEIN SWE
Settlement and Land Records Department
Ministry of Agriculture and Irrigation ( Myanmar)

Supervised by
DR. N.R. Patel
Scientist “SE”
Agriculture & Soil Division.
Indian Institute of Remote Sensing (IIRS) Dehradun, INDIA

CENTER FOR SPACE SCIENCE AND TECHNOLOGY


EDUCATIONIN ASIA AND THE PACIFIC (CSSTEA-AP)
AFFILIATED TO THE UNITED NATIONS

Indian Institute of Remote Sensing (IIRS)


4, Kalidas Road, Dehradun.
June, 2005

ii
CENTRE FOR SPACE SCIENCE AND TECHNOLOGY EDUCATION

IN ASIA AND THE PACIFIC (CSSTE-AP)

(Affiliated to the United Nations)

CERTIFICATE

This is to certify that Mr. U Thein Swe has carried out Pilot Project study entitled
“CROP YIELD ESTIMATION OF WHEAT BY INTEGRATING REMOTE
SENSING, LAND AND MANAGEMENT FACTORS” for the fulfillment of
Post Graduate Course in Remote Sensing and Geographic Information System of
CENTRE FOR SPACE SCIENCE AND TECHNOLOGY EDUCATION IN
ASIA AND THE PACIFIC (CSSTE-AP). This work has been carried out at Indian
Institute of Remote Sensing, Dehra Dun.

Supervisor

Dr. N.R. Patel


Agriculture & Soil Division
IIRS, Dehradun

Prof (Dr. Karl Harmsen) Dr. V. K.Dadhwal


Director, CSSTE-AP Dean, IIRS
Abstract

Many studies have revealed that there is correlation between remotely sensed
NDVI and yield. Few studies have applied remote sensing data at farmers’ field level to
estimate yield. At this scale agricultural production is a result of complex environmental
stresses including farmers’ management. This study, therefore, propose to investigate the
relationship between space-borne Satellite based NDVI and wheat yield at field level, and
combining NDVI with land and management factors for yield prediction at field level.

The study was carried out in Saharanpur district ( 29˚ 34’ 19” to 30˚ 23’ 58” N
latitude and 77˚ 07’ 24” to 77˚ 57’ 10” E longitude), Uttar Pradesh, India. High-resolution
LISS-III on board IRS-P6 satellite data for of IRS-P6, has been used for crop
discrimination and area estimation. Data was collected through interviewing farmers on
the management practices and farmer’s yield for rabi season (2004 – 2005). Crop yield
information was also gathered by actual crop harvest at randomly selected crop cutting
experiments (CCE). Soil types map (1:250000) from NBSS & LUP is digitized and used
to investigate relationship between NDVI, soil types and wheat yield. Average NDVI
pixel value of 3 × 3 window corresponding to CCE sites of wheat were extracted for
regression analysis with wheat yield. Land/soil productivity indices (LPI and Sys Index)
representing dominant pedons and corresponding to CCE sites are calculated and used in
regression analysis. Stepwise linear regression was used to relate yield,NDVI,land and
management factors to derive a yield estimation model. The results showed that there
exist a significant positive correlation (0.01% level) of field level wheat yield with NDVI,
land factors (LPI or Sys Index) and management inputs, particularly, number of irrigation
applied. It was found that space-borne NDVI alone explained 52.0 % (R2 adj.) of the yield
variability at field level. While, a combination of all the factors including NDVI
combined with land (LPI) and management (irrigation frequency) factors explained 66.3
% (R2 adj. ) of the yield variability. The study also showed that not all the factors affecting
yield also affect NDVI.

II
III
Acknowledgements

I wish to express my appreciation to Mr Cihat H.Basocak, GIS Officer of


UNESCAP, Bangkok for greeting me a scholarship to pursue a course of study at 9th Post
graduate course of CSSTE-AP, India Institute of Remote Sensing (IIRS), Director
General of Settlement and Land Records Department,Ministry of Agriculture and
Irrigation,Myanmar for allowing me to make use of the opportunity.

I am very thankful to Dr. V.K Dadhwal, Dean, IIRS, for his unrelenting
encouragement and effort towards providing all necessary facilities during the training
course.

My sincere and special thanks to Dr.N.R Patel, Agriculture and Soil Division,
IIRS, for his valuable guidance, encouragement advices and constructive criticism
throughout this paper. I wish to extend my sincere thanks to Dr. Suresh Kumar,
Agriculture and Soil Division, IIRS for his valuable comments, suggestions, help,
guidance and encouragement during the field study.

I would also like to express my sincere thanks to Dr. S.K Saha, Head, Agriculture
and Soil Division, IIRS for his rendering all necessary facilities for the training period
and project work.

I am also grateful to Dr. A. Velmurugan,Agriculture and Soil Division, IIRS,for


his valuable suggestions and helps for the project work.

I wish to thank my lab-mates for their help to my project work time.

Lastly, I wish to thank my CSSTE-AP officer trainees for their help, friendly
kindness during course period.

Dehradun
June, 2005 ( Thein Swe)
Myanmar

IV
Table of Contents
Abstract II

Acknowledgement III

Table of Contents IV

List of figures VII

List of tables IX

1 Introduction 1

1.1 The need for crop yield forecasting 2

2 Review of literature 4

3 Description of Study Area 7

3.1 Geographic setting. 7

3.1.1 Location and extent 7

3.1.2 Climate 8

3.1.3 Geology and soil 8

3.1.4 Agriculture and present land use 9

4 Materials and Methods 10

4.1 Materials 10

4.1.1 Remote sensing data 10

4.1.2 Ancillary data 10

V
4.1.3 Physiographic soil map 10

4.1.4 Land Management factors 10

4.1.5 Software used 10

4.1.6 Hardware used 11

4.2 Methods 11

4.2.1 Atmospheric and radiometric correction 11

4.2.2 Rectification 13

4.2.3 Digital image classification 13

4.2.4 Crop discrimination using high resolution data 13

4.2.5 Post classification 14

4.2.6 Spectral VI based statisticalyield models 14

4.2.7 Crop cutting experiments 14

4.2.8 Land Productivity Index (LPI) 15

4.2.9 Sys method of land evaluation 16

4.2.10 Single date yield estimation using high resolution satellite data 18

4.2.11 Multiple regression models 18

5. Result and Discussion 21

5.1 Atmospheric and radiometric correction 21

5.2 Crop inventory 22

5.3 Crop discrimination and acreage estimation using LISS-III data 23

VI
5..3.1 Spectral characteristics and spectral separability 23

5.3.2 Crop acreage estimation and accuracy assessment 25

5.4 Crop Yield Modeling 29

5.4.1 Spectral vegetation indices based yield estimation 29

5.4.1.1 NDVI Extraction 29

5.4.1.2 Land and management factors 30

5.5 Distribution of yield data 31

5.5.1 Yield prediction using NDVI 32

5.5.2 The effect of land parameters on yield and NDVI 35

5.5.2.1 Relationship between yield , NDVI and soil sub-group 35

5.5.3 The effect of management on yield 38

5.5.3.1 Urea Fertilizer Application 39

5.5.3.2 Irrigation Frequency 39

5.5.4 Correlation of RS, land and management factors 40

5.5.5 Single date image yield estimation using high resolution data 41

5.6 Discussion 45

5.7 Recommendation 46

6. Conclusion 47

References 48

Field Pictures 50

VII
List of figures

Figure 3.1 Location map of Saharanpur district 7

Figure 4.1 Crop cutting experiment in field 15

Figure 4.2 Flow diagram of crop acreage estimation 19

Figure 4.3 Schematic diagram for crop yield model development 20

Figure 5.1 Atmospheric correction of satellite data 21

Figure .2 Spectral reflectance of healthy vegetation 22

Figure 5.3 Spectral response curve IRS-P6-LISS-III 25

Figure 5.4 Land use/Land cover of Saharanpur district in 2004-05 26

Figure 5.5 Crop acreage estimation a) FCC, b) Classified image 28

Figure 5.6 a) Digital classified map, b) NDVI image 30

Figure 5.7 Histogram fitted with a normal probability curve 32

Figure 5.8 “Z” score of farmer’s yield 32

Figure 5.9 Regression analysis of NDVI and yield 33

Figure 5.10 Yield map 34

Figure 5.11 Relationship between soil type and yield 35

Figure 5.12 CCE sites on soil type map of Saharanpur district 36

Figure 5.13 CCE sites map and soil pedon location 37

VIII
Figure 5.14 Correlation of soil type, NDVI and yield 38

Figure 5.15 Effect of fertilizer application on yield 39

Figure 5.16 Relationship between irrigation frequency and yield 40

Figure 5.17 Correlation of Yield, NDVI, IRRI, LPI, SYS 40

Figure 5.18 a) Land use /land cover map b) Wheat mask map

c) wheat mask NDVI map d) Yield map 44

Figure 5.19 Correlation of farmer’s expected yield and observed yield 45

IX
List of tables

Table 4.1 Satellite data product 12

Table 5.1 Seperrability of different land use / land cover

classes of LISS-III data(Rabi) 24

Table 5.2 Land use/ land cover statistics of Saharanpur district

in 2004-2005 (Rabi) 26

Table 5.3 Error matrix showing the digital classification accuracy

of crops and other land use ( Rabi) 27

Table 5.4 Yield statistics of Saharanpur district in 2004-2005(rabi) 34

Table 5.5 Descriptive statistics of casual variables and its correlation

With wheat yield 42

Table 5.6 Models of Remote Sensing , Land and Management Factors 43

X
XI
1. Introduction

India underwent a series of successful agricultural revolutions, starting with the


"green" revolution in wheat and rice in the 1970s, the "white" revolution in milk and, in
the 1980s, the "yellow" revolution in oil seeds. Despite these major transformations, the
agricultural sector continues to be dominated by a large number of small landholders (70
% of rural people and 8 % of urban household depend on agriculture). The country is also
marked by large fluctuations in agricultural output, though to a declining extent with the
development of irrigation facilities, adoption of new technologies and changes in
cropping patterns (FAO, 2000a). The traditional approach of crop estimation in India
involves complete enumeration (except in a few states where sample surveys are
employed) for estimating crop acreage and sample surveys based on crop cutting
experiments (CCE) for estimating crop yield. The crop acreage and corresponding yield
estimate data are used to obtain production estimates.

These yield surveys are extensive; plot yield data being collected under complex
scientifically designed sampling design that is based on a stratified multistage random
sampling (Government of India 2002, Singh et al. 1992,). Final production estimates
based on this sampling method become available after the crops are actually harvested.
Although the approach is fairly comprehensive and reliable, there is a need to reduce the
cost and also to improve upon the accuracy and timeliness of crop production statistics.

Yield estimates predicted before actual production are required for taking various
policy decisions. Hence, early assessment of crop yield is necessary, particularly in
countries that depend on agriculture as their main source of economy. It enhances timely
provision of information for use in food security. In India, there is also a growing need for
micro-level planning and particularly the demand for crop insurance (Singh et al. 2002),
which increases the need for field level yield statistics.

At present, there is no model that relates field level yield to NDVI and no simple
method that produces quantitative pre-harvest data accurately and in time. With the
successful launching of satellites like IRS-1A and 1B in 1998 and 1991 respectively, and
other previous satellites, a lot of efforts are made to use remote sensing for yield
estimation.

1
1.1 The need of remote sensing for crop yield forecasting

Forecasting crop yield well before harvest is crucial especially in regions


characterised by climatic uncertainties. This enables planners and decision makers to
predict how much to import in case of shortfall or optionally, to export in case of surplus.
It also enables governments to put in place strategic contingency plans for redistribution
of food during times of famine. Therefore, monitoring of crop development, crop growth,
and early yield prediction are generally important.
Crop yield estimation in many countries are based on conventional techniques of
data collection for crop and yield estimation based on ground-based field visits and
reports. Such reports are often subjective, costly, time consuming and are prone to large
errors due to incomplete ground observations, leading to poor crop yield assessment and
crop area estimations (Reynolds et al.2000). In most countries the data become available
too late for appropriate actions to be taken to avert food shortage. In some countries
weather data are also used (de Wit & Boogaard 2001, Liu & Kogan 2002) and models
based on weather parameters have been developed. This approach is associated with a
number of problems including the spatial distribution of the weather station, incomplete
and/or unavailable timely weather data, and weather observations that do not adequately
represent the diversity of weather over the large areas where crops are grown (Dadhwall
& Ray 2000, de Wit &Boogaard 2001, Liu & Kogan 2002, Rugege 2002).Objective,
standardised and possibly cheaper/faster methods that can be used for crop growth
monitoring and early crop yield estimation are imperative.
Many empirical models have been developed to try and estimate yield before
harvesting. However, most of the methods demand data that are not easily available. The
models complexity, their data demand, and methods of analysis, render these models
unpractical, especially at field level. With the development of satellites, remote sensing
images provide access to spatial information at global scale; of features and phenomena
on earth on an almost real-time basis. They have the potential not only in identifying crop
classes but also of estimating crop yield (Mohd et al. 1994); they can identify and provide
information on spatial variability and permit more efficiency in field scouting (Schuler
2002). Remote sensing could therefore be used for crop growth monitoring and yield
estimation.

2
To achieve timely and accurate information on the status of crops and crop yield,
there is need to have an up-to-date crop monitoring system that provides accurate
information on yield estimates way before the harvesting period. The earlier and more
reliable information the greater the value (Hamar et al.1996, Reynolds et al. 2000).
Remote sensing data has the potential and the capacity to achieve this.

Keeping in view the potential of satellite remote sensing to quantitatively describe


actual crop conditions on remote wide area,non-destructive and /or real-time basis, the
present study was undertaken in Saharanpur district,(India) with following objectives :

 To discriminate crop types and wheat acreage using IRS-P6-LISS-III(3rd


March,2005) data during Rabi season.
 To investigate the relationship between NDVI and field level crop yield in
wheat.
 To investigate the relationship between wheat yield and NDVI combining with
land and management factors for yield prediction at field level

3
2 . Review of Literature

Various scientists in different part of world have demonstrated the use of remotely
sensed data for agricultural crop investigation. Agriculture is a major user of data from
satellite remote sensing. For more than a decade, in 1986 a project on Crop Acreage and
Production Estimation (CAPE) have been addressed on crop production estimates using
satellite observation in India which aimed at estimating production of crops viz, wheat,
rice, sorghum, cotton, groundnut and mustard in their major growing areas ( Navalgund et
al., 1991)

Recently, a FASAL (Forecasting Agricultural outputs using Space Agro-


meteorology and Land based Observations) is under operation which strengthen the
current capabilities from econometric and weather based techniques with remote sensing
application (Parihar, 1999)

Use of satellite remote sensing data for landuse / landcover inventory and
production forecasting in country like India presents a higher level of complexity due to
small field size, large number of crops, field to field variability in crop phenology and
management practices and presence extensive cloud cover during rainy season ( June to
September) ( Shahai, 1985). However, much progress has been made through intensive
studies on inventory and production forecasting for major crops using satellite data such
as Landsat MSS, IRS LISS-I, LISS-II and LISS- III ( Dubey et al.,1984, Saha et al., 1985,
Shahai & Dadhwal, 1990).

Development of crop yield models under IRS utilization Programme Uttar


Pradesh for wheat and paddy crops have shown that spectral vegetation indices at
maximum vegetative phase are highly related with final biomass of these crops.

Spectral vegetation indices, a linear or non-linear combination of reflectances


acquired in several bands are often correlated with biophysical attributes. Many studies
have experimentally or theoretically related vegetation indices to biophysical variables
( Asrar et.al.,Patel et. al., 1995,Wiegand et. al., 1992).

4
Recently, multi-spectral satellite data based indices along with agro-
meteorological indices were used for yield prediction for rabi paddy crop area of Nellore
district of Andhra Pradesh by using IRS-1A LISS –I data. Of the various spectral, agro-
meteorological yield models developed, they concluded that paddy yield estimation can
be improved by combining agro-meteorological indices like growing degree day ( GDD),
potential evapotranspiration (PET) with NDVI. (Saha and Jona, 1994)

Digital supervised classification of LANDSAT MSS data was used for


identification and district level acreage estimation of Kharif paddy ( Kalubrame, 1986).
The two stage stratified sampling approach and supervised digital classification of
LANDSAT MSS and TM and IRS-IA and IB, LISS –I data gave better estimates of
paddy crop acreage in larger areas such as a group of district or a state ( Parihar et.
al.,1987, Sharma et. al., 1990, Panigraphy et. al., 1991)

Krishna Rao et al., (1997) evaluated the feasibility of IRS –IC LISS-III data in
discriminating and estimation acreage of crops grown under multiple cropping situation
in 2mandals of Guntur district, Andhra Pradesh concluded that the data sat under
investigation has good potential to fulfill the objectives in multiple cropping situation.

For assessment of crop various and prediction of crop yield several vegetation
indices such as NDVI ( Thomas and Wehmanen, 1980),Greenness Vegetation Index
(GVI) Kauth and Thomas, 1976 and RVI (Pearson and Miller,1972) have been proposed.
In this approach the indices at particular growth stage of crop ( normally peak vegetative
growth/maximum LAI) are related to crop yield through regression techniques. The
integrated models incorporating satellite based vegetation indices, agro- meteorological
indices and time series trend predicted well the crop yield. Spectral profile modeling
approach using temporal satellite vegetation indices has also suggested for better
prediction of crop yield (Anonymous, 1995).

The timing of the image to be used for yield estimation is important. Though
Gielen et al. (2001)explained that there is good correlation between NDVI and yield but
using NDVI as an end-of season yield estimator gives unsatisfactory results because of
the problems of choosing the best time of the image to use, Muthy et al. (1994) found that
vegetation indices calculated from images taken at panicle initiation and heading stages

5
have high correlation with yield, therefore,they can best be used for yield prediction. The
findings of Muthy et al. (1994) agrees with the findings of this study.

However, it is difficult to have a single date image representing one phenological


stage at field level because of the differences in planting dates and the varieties used,
resulting in wide differences in crop phenological stages. To improve the predictability of
yield, Muthy et al. (1994) and Gat et al. (2000) proposed the use of time composited
multi-date images for yield prediction covering panicle initiation and heading stages and
considering maximum NDVI which normally occurs at heading stage. It is difficult,
especially in most tropical environments, to get a series of images due to clouds or other
logistical problems. In this case a single date image, as demonstrated, still provides good
information to predict end-of-season yield as long as it is within the time when there is
maximum vegetation (between panicle initiation and heading stage) and other production
factors are taken into account.

6
3. Description of the study area

The geographic setting of the study area, materials used during investigation and
methodology adopted to find out the desired objectives are briefly described below.
3.1 Geographic Setting

Location and extent, climate, geology, agricultural land use are delt in this section.
3.1.1 Location and Extent

The study area of Saharanpur district, Uttar Pradesh State is surround by


Dehradun district in the north, Yamuna river forms its boundary in the west which
separates it from Haryana district, in the east Haridwar district and in the south lies the
district of Muzaffarnagar. Saharanpur district is situated in north 29˚ 34’ 19” to 30˚ 23’
58”latitude and east 77˚ 07’ 24” to 77˚ 57’ 10”longitude.The area stretchs between
53G/1, 53G/2, 53G/5, 53G/6, 53G/9, 53G/10, 53G/13, 53G/14, 53F/8,53F/11, 53F/12,
53F/15 and 53F/16 topographical maps of 1: 50,000 scale which are prepared by Survey
of India, the study area is around 368,000 ha.

Dehradun District

Haryana
District

Saharanpur

Fig. 3.1 : Location map of study area

7
3.1.2 : Climate

The climate of study area is the average climate of Uttar Pradesh in general but its
northern position and its proximity to the hills give its on peculiarity. Though the region
lies well outside the tropic yet its climate like that of the rest of north India is essentially
tropical because of Himalayan chain. It belongs to the uppermost part of the upper Ganga
plain which is a sub-humid region between the dry Punjab plain and the humid middle
Ganga plain within the monsoon region of the great plains and naturally partakes the
characteristics of the to adjoining regions.

The average temperature recorded is 23.3 degree centigrade June being the hottest
month while January is the coldest one. The highest percentage of humidity i.e. 72 to 85
% is found during the rainy season at the lower range of humidity between 29 to 51.5 %
is recorded in the summers. The eastern part of the region is more humid than the western
part and relative humidity tends to increase in the winters season. Pressure of the region is
inversely related to the temperature July recording the lowest while December recording
the highest pressure. The average pressure of the district is found to be around 979 lbs.
The average annual rainfall of Saharanpur district is found to be around 650mm.The
highest rainfall is July and August and the lowest is in March and April.

3.1.3 Geology and soil

The Saharanpur district is apart of the Indo-Gangetic alluvial plain with the
alluvium belonging to the Pleistocene as well as sub-recent and recent time. Saharanpur
forms the most northerly position of the Doab land which stretches between the holy
rivers of the Ganges and Yamuna. The Shivalik hills rise above it on the northern frontier.
In Saharanpur district, many verities of land forms are lying, such as : summit and ridge
of Shivalik hill, upper piedmont, lower piedmont, fluvial plain, old alluvial plain, recent
alluvial plain and active flood plain. Mainly, old alluvial plain is lying more than half of
the whole district. The main characteristics of the district can be divided into four parts:

(1) Shivalik hill tract

8
(2) The Bhabar land
(3) Bangar land
(4) Khadar land ( Yamuna, Hindon)

3.1.4 Agriculture and Present Land use

Saharanpur is primarily agricultural district. Roughly 70% of the land is used for
agriculture. Agriculture plays an important role in the economy of the district. Hence,
major agricultural systems viz, paddy, wheat, sugarcane and orchards are practiced in the
district. The developed and fertility alluvial plain of Saharanpur district is contributed by
the network of eastern Yamuna canal and its distributaries of many channels. The eastern
Yamuna canal runs through the center of district from north to south. One significant
feature is that even thought the agricultural land for food crops has reduced in recent
years the food production has increased considerably. The significance of commercial
crops have increased manifold as a consequence of sugarcane production. In study area,
the commercial and important crops of region are paddy, wheat, sugarcane, orchards and
mustard.

9
4. Materials and Methods

4.1 Materials
Data used for the study and software specification are delt in this section;

4.1.1 Remote Sensing Data

Satellite Sensor Product Path/Row Date of Acquisition Source

IRS-P6 LISS-III Hard Copy 96 /49 3-03-05 NRSA

4.1.2 Ancillary Data

Survey of India Toposheet

Toposheet No’s : 53G/1, 53G/2, 53G/5, 53G/6, 53G/9, 53G/10, 53G/13,

53G/14, 53F/8, 53F/11, 53F/12, 53F/15, 53F/16

Scale : 1 : 50,000

Land sat ETM : 30 Meter Resolution

4.1.3 Physiographic Soil Map

Soil type Map of Saharanpur district (1 : 250000) from NBSS & LUP

4.1.4 Land Management Factors

Land management factors was collected for yield from farmers by interview
method and Crop yield information was also gathered by actual crop harvest at randomly
selected crop cutting experiments (CCE).

4.1.5 Software Used

1) For image processing

a) ERDAS Imaging 8.7

10
2) Data generation

a) ERDAS imaging 8.7


b) ARC GIS
3) For GIS analysis

a) ARC VIEW
b) ARC GIS
c) ILWIS 3.2

4) For Calculation and report writing

a) MS Office
b) MS Excel

4.1.6 Hardware Used :

Pentium III,128 Mhz memory,

4.2 Methods

The methods used during the investigation is briefed in the following section.

4.2.1 Atmospheric and radiometric correction

The reflectance of various land use / land cover features reaching the satellite
sensor is attenuated by atmosphere. In order to make the top of the atmosphere
reflectance match the surface reflectance, atmospheric correction has to be carried out. In
the present study improved dark object subtraction (COST method) proposed by Chavez
(1996) was used to correct the image for atmospheric effects. The methodology is
described as follows.

The model first converts each minimum DN value to an at-satellite minimum


spectral radiance value. The radiance image is calculated by the following equation.

Lλ ,min = LMINλ + QCAL * (LMAXλ - LMINλ ) / QCALMAX

Where, QCAL is minimum digital number (DN),

QCALMAX = 255, and LMINλ and LMAXλ are constants described in the table 5.1.
These constants are given in the header file of the satellite data product. The Earth-Sun

11
Distance, sun elevation angle and minimum DN values are the other required inputs. For
each band, the theoretical radiance of a dark object is assumed to have a reflectance of
one per cent (Moran et al. 1992 and Chavez, 1996) and calculated using the following
equation.

Lλ , 1% = 0.01 * d2 * cos2 θ / (π * ESUNλ )

Where, ESUNλ = mean solar exo-atmospheric spectral irradiance (table 4.1)

d is the sun-earth distance and θ is the solar zenith angle (90-solar elevation angle).

Haze correction is computed from the dark object values (Chavez 1996):

Lλ ,haze = Lλ ,min - Lλ ,1%

The radiance image is then converted into reflectance by the fundamental radiance to
reflectance (rho) equation:

ρ = π * d2 * (Lλ sat - Lλ haze ) / ESUNλ * cos2 θ

The output from the model is in reflectance units, which should range from 0 to 1.
Occasionally, values greater than 1 are obtained, and these generally correspond to bright
objects (e.g., clouds, snow) or sometimes noise (randomly or systematically scattered
saturated pixels).

Table 4.1 Gain, offset and (mω cm-2 str-1 µ m-1) and band pass solar exo- atmospheric
irradiance,E0 (mω cm-2 str-1 µ m-1)) for the satellite data used in the study

Table 4.1 satellite data product

Satellite Sensors Bands Gain Offset E0


Green 12.06 0.0 184.892
Red 15.13 0.0 157.679
IRS – P6# LISS – III
NIR 15.76 0.0 109.338
SWIR 3.397 0.0 23.786

12
*
and # Pandya et al., 2003 and 2004, respectively.

4.2.2 Rectification

A full scene (path/row:96/49) of high-resolution satellite data from LISS – III


sensor onboard IRS 1D and IRS P6 were georeferenced in UTM projection using ground
control points (GCPs) from the Survey of India topographical maps at 1:50,000 scale.
These georeferenced images were then resample to 23.5m pixel size using nearest
neighbour technique and the images were clipped using the study area boundary mask.

4.2.3 Digital Image Classification

Before final classification of satellite data spectral seperability between


crop and other land use/ land cover classes were evaluated multi band scatter diagram of
training classes. The crops discrimination of study area was generated by digital
supervised classification following Maximum Likelihood Classification (MXL)
algorithm. The acreage of crops and other land use/ land cover were estimated from the
MXL report.

4. 2. 4 Crop discrimination using high resolution data

To discriminate wheat crop from various land use and land cover classes, crop
inventory was carried out using the satellite data. Ground truth was collected by
integrated use of Global Positioning System (GPS), geo-coded temporal FCC of MODIS
and SOI toposheets (1:250,000 scale). After critically examining the spectral variations in
the geo-coded FCC, field visits were planned at various phonological stages of wheat
using complete enumeration technique. The training areas (samples) for different land use
and crop classes (wheat, sugarcane, orchard, forest, fallow etc) were defined interactively
on satellite image based on the field visits. For crop identification, a supervised
maximum-likelihood classification (MXL) was performed using all optical bands for IRS
P6 LISS – III data.

13
Accuracy assessment of classified pixels was done using independent reference
sites of the study area, collected using GPS. Overall accuracy was defined as the
percentage of total independent reference pixels that were correctly classified by the
MXL algorithm. Producer’s accuracy was calculated by dividing the number of pixels
correctly classified for each crop by the total number of independent reference pixels for
that crop, while user’s accuracy was the number of correctly classified pixels divided by
the total number of classified pixels for that crop. Kappa coefficient was calculated to
measure the significance of classification results relative to chance agreement. A kappa
value of zero indicates that the classification is no better than random assignment of
pixels, while a value of one indicates perfect agreement between training pixels and their
prescribed classes (Lillesand and Kiefer, 2000).

4.2.5 Post Classification sorting

After classification with MXL, some classification errors could be already detected
during a visual examination of classified image. Reclassification was done by merging
relevant classes and generation or smoothing of classified image was done by using
statistical filtering with 3 x 3 mode filter to improve the accuracy.

4.2.6 Spectral VI based statistical yield models

Timely and reliable crop yield prediction is an essential component of crop


production forecasting system. The use of satellite remote sensing data for operational
yield assessment and forecasting at regional scale is of great significance in providing
food security and making timely policy decisions. Operational yield forecasting using
remotely sensed satellite data has generally followed empirical regression approach
making use of high and coarse resolution and single and multiple date satellite data
products.

4.2.7 Crop cutting experiments

The crop cutting experiments (CCE) were carried out at maturity stage of wheat in
randomly selected plots (n = 44) in the study area. Wheat area of 1m x 1m size
representing the sample site was harvested and its location was recorded with GPS. Grain

14
yield and biomass was separated and oven dried to obtain final grain yield for different
sample sites.

Fig –4.1 Crop Cutting Experiments in Field

4.2.8 Land Productivity Index ( LPI)

LPI is based on general characteristics of the soil profile, texture of the surface
soil, soil of the land, climate and other physical factors affecting use of land. It is a
parametric approach of assessing the potential of land. In this approach a numerical value,
which in principal, ranges between 100 ( for non- constraint situation) and 0.0 (for a
severe constraint situation)is assigned to each characteristics. The Value given to each
land quality (characteristics) A,B,C,X and Y are multiplied, and the final product is
reconverted into land productivity class with fixed index values.

Land Productivity Index ( LPI) = A*B*C*X*Y

Where factors are decimal equivalent of percentage rating.

A = General characteristics of soil profile

B = Texture of the surface soil

C = Slope of the land

X = Miscellaneous factors; reaction of surface soil, fertility, erosion

Y = Average annual rainfall

15
Land productivity classes

Classes Ranges
Excellent (Class I) 80 – 100
Good (Class II ) 60 – 80
Fairly Good (Class III) 40 – 60
Average ( Class IV ) 20- 40
Poor ( Class V ) 10 – 20
Very Poor ( Class VI ) < 10

4.2.9 SYS method of land evaluation

It is a parametric approach of FAO frame work of land evaluation in which


numeral rating of the different limitation levels of the land characteristics in a numeral
scale from maximum (normally 100) to a minimum values is assigned.

In the study, Typical pedons of the map units were assessed for its suitability to
wheat crop. Important soil and land characteristics were considered to reduce repetition of
related characteristics. These characteristics, such as drainage, texture and coarse
fragments, soil PH, EC, etc, were considered for the evaluation. The Important
characteristics rated in wide scale (100-25), and a less important characteristics in
narrower scale (100-60). This introduces the concept of weighting factor. The rooting
depth of land utilization types (LUTs), supposed to be most important for assessing
suitability. Thus, weighted average of soil characteristics of horizon-wise of each pedon
were computed considering the 125 cm rooting depth of wheat crop. Weighting factors
were assigned to each horizon for the computation. Thereafter, the characteristics were
rated in numerical scale (100-25/60). These individual ratings were used to calculate Sys
Index as :

16
Sys Index = A * B/100 * C/100* ……….

( A, B and C are ratings of soil and land characteristics)

The successful application of the system applies the respect of the following rules:

1. The number of land characteristics to consider has to be reduced to a


strict minimum to avoid repetition of related characteristics in the
formula, leading to depression of the land index.
2. An important characteristics is rated in a wide scale ( 100 – 25), a less
important characteristics in a narrower scale ( 100 – 60). This
introduces the concept of weighting factor.
3. The depth to which the land index has to be calculated must be defined
for each land utilization type.

The depth to be considered should coincide with the normal depth of


the root system in a deep soil.

Number of sections and weighting factors for different depths

No. of equal
Depth (cm) Weighting factors
sections
125 – 150 6 2.0, 1.5, 1.0, 0.75, 0.5, 0.25
100 – 125 5 1.75, 1.5, 1.0, 0.5, 0.25
75 – 100 4 1.75, 1.25, 0.75, 0.25
50 – 75 3 1.5, 1.0, 0.5
25 – 50 2 1.25, 0.75
0 – 25 1 1.0

The rooting depth of LUTs considered in the evaluation were :

Sugarcane - 150 cm

Wheat - 125 cm

Weighting factor for rooting depth of each LUTs were considered to computing weighted
value of land characteristics. Sys index were obtained for each typical pedons of soilscape
units to assess the suitability. The Sys index (land index) computed for each pedon were

17
classified into four suitability classes of S1 ( >60), S2 ( 40 – 60 ), S3 (20–40), and N ( <
20).

4.2.10 Single date yield estimation using high-resolution


satellite data

The supervised classification images of IRS P-6 LISS – III was used to mask the
respective satellite reflectance to generate reflectance images for wheat and the NDVI for
wheat was generated. The harvesting sites identified during CCE were overlaid on the
wheat mask image and an average NDVI of 3 x 3 pixel window was extracted. Empirical
relationships were developed between the yield obtained from the CCE and
corresponding average NDVI to develop single date yield prediction model for wheat.

4.2.11 Multiple regression model

Remotely sensed NDVI in combination with land (LPI) and management (irrigation
frequency/ factors information) corresponding to CCE sites were used in causal variables
in multiple regression analysis. Different kinds of multiple regression models involving
combination of NDVI, LPI, Sys and Irrigation frequency were developed. The goodness
of fit or performance of models were evaluated based on coefficient of determination of
( R2 ) and standard error of estimates (SEE).

18
Fig – 4.2 Flow diagram for Crop acreage estimation

19
SOI Toposheet Satellite Data
1 : 50000 IRS-P6-LISS-III

Georeferencing to
Image Rectification
Scanned Image

Digitization of Subset Study


Study Area Area

District Boundary
Image

Ground Truth Training Signature


Training Samples Generation

Supervised Classification
(MXL)

Accuracy Assessment

Crop inventory

20
Fig- 4.3 Schematic diagram for crop yield model development

Satellite Image NDVI Image GPS Location


Sample Sites

Training Siganture Wheat Crop


Generation Masked NDVI Management
Factors- Irrigation,
Fertilizer

Classified Image Extraction of Mean 3x3 Pixel


Window Pertaining to Sample
Sites
Land Factors
Wheat Crop -LPI,Sys
Masked Empirical yield
model development

Crop Cutting
Experiments(CCE
Validation
)

21
5.0 : Result and Discussion

5.1 Atmospheric and radiometric correction

The satellite images have to be corrected for the sensor anomalies and atmospheric
effects before it is used for any further analysis. Radiometric correction was carried out
using the gain and offset values of the satellite sensor present in the header file of the
image. Later this radiance image was converted into at surface reflectance image by
correcting the image for atmospheric effects (chavez, 1996). The result of atmospheric
correction is depicted using the histogram corresponding to the difference between the
NDVI obtained after atmospheric correction (NDVIchavez) and the NDVI obtained without
atmospheric correction (figure-5.1). It can be observed from the figure that, there is a
positive shift in the graph that is centered on 0.15 NDVI. This concludes that,
atmospheric correction has increased the satellite reflectance by the order of 8 per cent
which otherwise would have not been observed due to the attenuations caused by
atmospheric effects. These results are in conformity with the results obtained from
Sobrino et al., 2004.

12

10

8
pixels (10
3
)

0
-0.2 -0.1 0.0 0.1 0.2 0.4
NDVI corr -NDVI toa

Fig -5.1 Atmospheric correction of satellite data

22
5.2 Crop inventory

Green plants have a unique spectral reflectance influenced by their structure and
composition. The proportion of radiation reflected in different parts of the spectrum
depends on the state, structure and composition of the plant. In general, healthy plants and
dense canopies, will reflect more radiation especially in the near infrared region of the
spectrum. In the visible part of the spectrum (0.4 μm – 0.7 μm), plants absorb light in the
blue (0.45 μm)and red (0.6 μm) regions and reflect relatively more in the green portion of
the spectrum due to the presence of chlorophyll. High photosynthetic activity will result
in lower reflectance in the red region and high reflectance in infrared region of the
spectrum. In cases where plants are subjected to moisture stress or other conditions that
hinder growth, the chlorophyll production will decrease, This in turn leads to less
absorption in the blue and red bands (Dadhwal & Ray 2000, de Wit & Boogaard 2001,
Janssen & Huurneman 2001, Woldu 1997). In the near-infrared portion of the spectrum
(0.7 – 2.5 μm), green plants reflectance increases to 40 – 60%. Beyond 1.3 μm, there are
dips in the reflectance curve due to absorption by water in the leaves, more free water
result in less reflectance. Figure 5.2 shows an ideal reflectance curve from healthy
vegetation.

Fig - 5.2 Spectral reflectance of healthy vegetation

5.3 Crop discrimination and acreage estimation using

23
IRS-P6-LISS- III data

5.3.1 Spectral characteristics and spectral seperability

Using the field wise ground truth collected with GPS, a large number of fields and
other land use classes were marked in LISS – III image acquired in march, 2005. And, it
could be easy to identify wheat training sets in view of large contiguous areas and best
suitable satellite data acquisition date for wheat crop. It was also feasible to identify
standing sugarcane ( i.e maturity stage), however greater difficulty was noticed in
identifying training areas pertains to newly planted sugarcane. This crop with poor
canopy cover limit the possibility of being separate out from fallow land in the image.

The spectral seperability based on Transformed Divergence ( TD ) for the major


rabi crops and other land use training classes using LISS-III ( band –3,2,1) data were
presented in ( Table-5.1 ). In general larger transformed divergence, greater the statistical
distance training patterns and the higher the probability of correct classes. The maximum
possible divergence value is 2000. As general rule, if the value of the difference between
two classes are > 1900, the two classes can be well separated. Between 1700 and 1900, it
is weak. Below 1700 the two classes cannot be separated (Swain and Davis, 1978).
Results revealed that different pairs of combination of wheat, sugarcane, orchards, fallow
land, water body, settlement etc., except orchards-fallow land(with water), forest-fallow
land, forest-orchards, riverine forest-orchards, plantation forest-fallow land and riverine
forest –fallow land were having >1900.These pairs can not well separate. Hence, it could
be inferred that except these pairs, other crop and land use are highly separable among
them. The best average spectral seperability in terms of TD for all pair wise combinations
of crop and other land use training classes was 1944.

24
TABLE - 5.1 Seperability of different land cover classes of LISS – III data(Rabi)

Class_
Fallow Riverine Plantation River Water
Name Wheat Sugarcane Orchard Forest Settlement
land Forest Forest Bed Body

Wheat 0 1999 2000 2000 2000 2000 2000 2000 2000 2000

Sugarcane 2000 0 1967 1992 1999 1995 1908 1985 2000 1985

Orchard 2000 1967 0 1591 1283 1653 1995 1978 2000 2000

Fallow_land 2000 1992 1591 0 1533 1991 2000 2000 2000 2000

Forest 2000 1999 1283 1533 0 1964 2000 2000 2000 1999

Riverine Forest 2000 1995 1653 1991 1964 0 1995 1895 2000 1999

Plantatio Forest 2000 1908 1995 2000 2000 1995 0 1837 2000 2000

Settlement 2000 1999 1978 2000 2000 1895 1837 0 1918 1997

River Bed 2000 2000 2000 2000 2000 2000 2000 1918 0 2000

Water Body 2000 1985 2000 2000 1999 1999 2000 1997 2000 0

*Best average seperability over all pair wise combination of signature: 1944

25
Spectral Response Curve (LISS_ III)

16 0
14 0

12 0
( DN Value)

10 0
80
60
40

20
0
Band_1

Band_3
Band_2

Band_4
Wheat Sugarcane Orchard Fallow_land
forest Settlement water_body

Fig - 5.3 Spectral response curve IRS-P-6 LISS-III

5.3.2 Crop acreage estimation and accuracy assessment

The major crop discriminated were wheat and sugarcane with other land use
classes. The distribution of area under crops and other land use along with false color
composite is presented in are Table and Figure. The wheat is extended 187298 ha, more
than half percent of total geographical area. In time of acquiring satellite data, some
fallow lands are irrigated for new plantation of sugarcane. Likewise, some orchards are
also irrigated. Due to this condition, orchards and fallow land are confused for
classification. The accuracy assessment between the training sites and classified pixels in
the out image was generated through error matrix. The error matrix with computed
omission and commission error. Kappa coefficient and overall classification accuracy has
been presented in Table 3. The over all accuracy only consider the diagonal element
includes the omission and commission error while Kappa coefficient includes off
diagonal elements (omission and commission error) as a product of row and column and
is a better technique for evaluation of classification accuracy (Congalton, 1994).

26
The overall accuracy for all the crops and other land use classes is more than 80%
and Kappa coefficient is 0.91.The accuracy achieved is much above the acceptable
accuracy (80%) for any kind of thematic map. The wheat crop in the study area shown
more than 99% accuracy.

Land Use Area (ha) Area %

Wheat 187298.00 50.89 %


Sugarcane 22200.10 6.03 %
Orchard 50411.40 13.70%
Fallow land 34057.80 9.25%
Forest 28772.70 7.82%
Riverine forest 2635.83 0.72%
Plantation forest 2051.39 0.56%
Settlement 28602.60 7.77%
River bed 9479.76 2.58%
Water body 2508.82 0.68%
Total Area (ha) 368018.40

Table-5.2 Land use / land cover statistics of Saharanpur district in 2004-2005 (Rabi)

Land Use Statistics of Saharanpur District


(2004,-2005,Rabi)
2051.39 ha 9479.76 ha
28602.60 ha 2508.82 ha
2635.83 ha
28772.70 ha
187298.00 ha
34057.80 ha

50411.40 ha

22200.10 ha

Wheat Sugarcane Orchard Fallow_land


Forest Riverine_forest Plantation_forest Settlement
River_bed Wat er_body

Fig –5.4 Land use statistics of Saharanpur district in 2004-2005 Rabi season

27
TABLE –5. 3 Error matrix showing the digital classification accuracy of crops and other land use (IRS –P6-LISS_III)data (Rabi)

Class_ Sugar-
Fallow Riverine P;antation Settle- River- Water Accuracy
Name Wheat cane Orchard Forest Total %
land Forest Forest ment Bed Body
Wheat 2047 0 0 8 0 0 0 0 0 1 2056 99.56
Sugarcane 19 254 1 0 0 1 12 9 0 3 299 85.62
Orchard 0 0 198 6 16 12 0 9 0 3 244 81.15
Fallow_land 0 0 6 651 36 0 0 0 0 0 693 93.94
Forest 0 0 13 7 793 0 0 0 0 0 813 97.54
Riverine_Forest 0 0 3 1 0 792 0 41 0 4 841 94.17
Plantatio Forest 0 7 0 0 0 0 493 43 0 1 544 90.63
Settlement 0 0 1 6 0 18 7 1628 66 0 1726 94.32
River_Bed 0 0 0 0 0 0 0 31 758 0 789 96.07
Water_Body 0 0 0 0 0 1 0 0 0 429 430 99.77
Total 2066 261 222 679 845 824 512 1761 824 441 8435 93.28
Accuracy% 99.08 97.32 89.19 95.88 93.85 96.12 96.29 92.45 91.99 97.28 94.94 95.49

Confusion Matrix
Average User Accuracy = 93.28 %
Average Producer Accuracy = 94.94 %
Overall Accuracy = 95.49%
Kappa Statistics = 0.91

28
(a) (b)

Fig – 5. 5 Crop Acreage Estimation (a ) False Color Composite ( IRS-P-6 ,LISS-III -Rabi) (b) Digitally Classified Image

29
5.4 Crop Yield Modeling

Crop yield prediction models are necessary for assessing the production of
particular crop in region. Most of these models use either agronomic variables or
meteorological variable or combination of both and they become highly location specific.
In this study, spectral data is an integrated with land and management factors.

5.4.1 Spectral vegetation indices based yield estimation

Crop yield is key element for rural development and an indicator of global food
security. As global food demand continues to grow, crop yield assessments on a regional
scale will be increasingly important. In the present study empirical models which directly
relate single-spectral satellite data or derived parameters (Vegetation Indices, VIs) to crop
yield was used in yield estimation of wheat. In this approach, NDVI at particular growth
stage (normally, maximum vegetation growth) is related to final crop yield through
regression techniques and pre-harvest crop yield is predicted with the assumption that
crop growth is not affected after that growth stage.

5.4.1.1 NDVI Extraction

The IRS – P6 – LISS-III image with a resolution of 23m taken on 3rd March 2005 was
based on the analysis of the information given by the farmers on the date of transplanting
and date of harvesting which gave the general crop calendar for the 2004 – 2005 Rabi
crop. The image was processed in a GIS environment and an NDVI map was generated
using band 3 (NIR) and band 2 (red) of the IRS image by applying the formula:

NDVI = (Band 3 – Band 2) / (Band 3 + Band 2)

Wheat crop NDVI is generated from NDVI satellite image and wheat masked
image. The first step in the analysis was to establish the relationship between NDVIs and
yield data at field level. Secondly, land and management parameters were compared with
yields and NDVIs. Then through regression the relationship were further tested to
establish yield model.

30
Data on land and management practices and respective yield data were collected
from farmers through interviews. Data units as reported by farmers were converted into
standard metric (S.I.) units. The IRS satellite image of 3rd Mar 2005 provided the field
level NDVI data. The total sample size for this study consisted of 44 valid fields.

Parametric statistical analysis techniques require data to be distributed normally.


Means and standard deviations are useful to describe data but become poor when the data
are not normally distributed. Histograms, stem-and-leaf plots and box plots can also be
used to visualize data. They help to show their distribution characteristics.

(a) (b)

Fig 5.6 : ( a )Digital classified Map ( b ) NDVI Image

5.4.2 Land and Management Factors

Vegetation density is the most obvious physical representation of subsequent yield


from crops. The density and health can be monitored using remotely sensed images that
measure chlorophyll activity and vegetation vigor. The spectral reflectance is a

31
manifestation of all important factors affecting the agricultural crop and cumulative
environmental impacts on crop growth (Liu & Kogan 2002, Singh et al. 2002), therefore
remotely sensed data could be used to monitor crop condition through NDVI.

Management practices in the production system and how land is utilized will have
an effect on the overall productivity. In this respect, crop growth and crop yield is a
response to the type of management and the quality of the land unit.

Based on the above, hypothesis adopted in this study are as follows:

1. There is significant relationship between NDVI, and yield at field level.

Yield = ƒ ( NDVI )

2. There is significant relationship between NDVI, field level management and land.

NDVI = ƒ (Land, Management)

3. Single date multi-spectral images can be used to predict yield at field level.

Integrating hypothesis 1 and 2 leads to:

Yield = ƒ (NDVI, Land, Management)

Where only those land and management factors remain relevant that have no
impact/relation with NDVI, the rest are now no longer relevant.

5.5 Distribution of Yield Data

The yields from the Crop Cutting Experiments was from 10.55 to
53.13 Q/ha with a mean of 33.05 Q/ha and median of 32.72 Q/ha.
Figure (5.7)- shows a histogram fitted with a normal curve and figure
(5.8) shows Z-scores of the 44 yield data. The data seem to follow a
normal distribution. Testing for normality by the two tailed Kolmogorov-
Smirnov test gave a (p) value of 0.86 which confirms the hypothesis
that the data is normally distributed, therefore parametric analysis

32
techniques can be employed for further analysis without fulfilling any
transformation requirement.

10
Sd = 10.187
Mean = 33.05
Kolmogorov_Smirnov “Z” test=0.815
8
Frequency

0
10 15 20 25 30 35 40 45 50

YIELD(Q/ha)

Fig – 5. 7 Histogram fitted with a normal probability curve

2.5
2 Sd = 10.187
1.5 Mean = 33.05
Expected normal

1 Kolmogorov_Smirnov “Z” test=0.815


0.5
0
-0.5
-1
-1.5
-2
-2.5
0 10 20 30 40 50 60
Observed value

Fig- 5. 8 Z Score of farmer’s yield

5.5.1 Yield prediction using NDVI

The crop yield models were developed by regression analysis using normalized
difference vegetation index (NDVI) at maximum growth stage of wheat and measured
wheat yields from crop cutting experiments (CCE). The result from the regression

33
analysis for yield estimation is presented in figures5.10. This study established that there
is a significant positive relationship between remotely sensed NDVI and CCE based yield
(Adj. r2 = 0.521), where production is dependent on many factors acting upon crop
growth. This clearly shows the potential of using NDVI for regional yield prediction for
wheat.

60
Yield (Q/ha) = 60.84*NDVI – 9.895
50 (Adj. R2 = 0.521, SEE = 7.142 N = 44)

40
yield (Q/ha)

30

20

10

0
0.2 0.4 0.6 0.8 1
NDVI

Fig 5.9 Regression analysis of NDVI and Yield

This study established that there is a significant positive relationship


between remotely sensed Normalised Difference Vegetation Index and field level yield (r
= 0.521, p = 0.001, standard error estimation (SEE)=7.142) where, production is at the
mercy of many factors acting on the crop. This analysis clearly shows the potential of
using NDVI for yield prediction at field level. NDVI explains the yield variability caused
by date of transplanting, soil, and other management factors.The result of this analysis is
a significant relationship between NDVI and field level yield. High NDVI value is related
to high yield. The result of yield is also reasonable with farmer’s expected yield.
Therefore, this regression analysis using NDVI is significant relationship to field level
yield prediction.

34
Fig –5.10 Yield map

Yield (Q/ha) Area(ha) Area %

<25 22526.26 12.03


25 - 30 20956.56 11.19
30 - 35 38590.51 20.63
35 - 40 63357.76 33.83
40 - 45 40380.30 21.56
>45 1486.60 0.79
Total 187297.99

35
Table 5.4 Yield statistics of Saharanpur district in 2004-2005 (Rabi)

5.5.2 The effect of land parameters on yield and NDVI

5.5.2.1 Relationship between yield, NDVI and soil sub-groups

Soil is a major role in crop production. It is a medium for water and nutrient
supply to crops. Its natural characteristics determine the availability and supply of these
resources to the crop. Fig 5.11. shows the distribution of yield by soil sub-group. The box
plots (figure-5.11 ) indicate that the highest yield in soil sub group AP-FL. Most of the
sample sites were in soil sub-group AP_ Fl, and the least samples in sub-group AP_FS.
This bias in sampling frequency relates to extent each subgroup occur.

In study area, the highest yield is found in AP_FL(Alluvial Plain, Fine Loamy)and
the lowest yield is found in UP_CL(Upper Piedmont, Coarse Loamy)of subsoil group.
Mostly, high vigor wheat crop and high NDVI value are also found in AP_FL soil
subgroup.

60

50

40
yield (Q/ha)

30

20

10
0
N= 3 5 22 2 9 3
AF

AP

LP

UP
P_

P_

_F
P_

_C

_C
FL

FS

L
SS

SOILTY PE

36
Fig- 5.11 Relationship between soil type and yield

The box plots, figure 5.11, show the distribution of yield in different soil sub-
groups. The box plots suggest more variation in soil sub-group AP_FS and least in sub-
group AFP_SS. Testing for differences in mean yield by soils suggested that at least one
soil sub-group is significantly different from other soil sub-groups ( p = 0.001).

A step-wise forward regression analysis with all soil sub-groups showed that
yields from soil UP-CL are significantly different yields from other soil sub-groups.
These results suggest that soil has a significant impact on growth and condition of wheat,
which can be measured through remotely sensed NDVI.

Fig – 5.12 CCE Sites on Soil Type map of Saharanpur District

37
Fig – 5.13 CCE sites map and Soil Pedons location

In this study found NDVI and soil type are highly correlated to yield. Fig-5.14 :
showed, high NDVI vale is directly related to high yield. Likewise, the soil sub-group of
AP-FL and AP-FS is more yield than other soil sub-group. NDVI and soil type are also
highly correlated. If soil is more fertile, the crop vigor will be more developed. So, NDVI
value is depend on the vigor of healthy crop.

38
Fig – 5. 14 Correlation of NDVI, soil type and yield

5.5.3 The effect of Management on yield

Farmers apply different management practices to outcome reasonable yields from


their farms and to get more income from agricultural activity. They operate at different
technological levels according their experiences and knowledge. Data on operation
sequence that was followed by individual farmers was collected and analyzed. The data
included all levels of technological operations and production level practiced in the area.
Management operations as reported by farmers were based on the date of the image. All
operations before this date were analyzed to find their effect on the final yield and NDVI.

39
5.5.3.1 Urea fertilizer applications.

Figure 5.15 : a box plot showing the effect of urea fertilizer application frequency
on yield. The box plot suggests more yield if fertilizer was applied more. Analysis is used
data on collected from field of 14 samples. Fertilizer application in analysis is from(n=4)
180 Kg to (n=5) 360 Kg /ha. Mean fertilizer usage in wheat crop is 276 kg/ha. Testing to
find if there is a significant difference in yield, Study suggested that the number of
fertilizer applications relates to the yield variability between fields.

50
Y ield (Q/ha)

40

30

20
N= 4 5 5
180.00 270.00 360.00

UREA (kg/ha)

Fig- 5.15 Effect of fertilizer application frequency on Yield

5.5.3.2 Irrigation Frequency

Irrigation water supply is a major production for wheat in rabi season. The power
problem as expressed by farmers affected the supply of water in most of the fields at
different times of plantation. In wheat sowing time of rabi, rain is very less. Most of
farmers used irrigation pump supplying. Mostly,wheat cultivation were used from eastern
Yamuna canals and its distributary’s channels. In this study, irrigation frequency (n=28)

40
was found mini (1) and maxi(4)times. The correlation of irrigation frequency and wheat
yield was high significant in determination (0.716). Fig 5.16 shows the mean yield and
irrigation frequency for the water regimes as expressed by farmers. Box plots show the
frequency of water (n=4)was high distribution on yield than less frequency (n). Analysis
found irrigation frequency are highly related to yield.

Yield = 11.85 + 8.13*Irrigation applied


50.00 R2 = 0.51
Error bars (95% CL of Mean)
yield (Q/ha)

40.00


30.00 n= 11
n= 4
n= 10

20.00

10.00
n= 3

0.00
1 2 3 4
Irrigationfrequency

Fig- 5.16 Relationship between Irrigation frequency and Yield

5.5.4 Correlation of RS, Land and Management Factors

41
Fig–5.17 Correlation of
60
YIELD
40
20
0
.8
NDVI

.6
.4
3
IRRI

2
1
80
LPI

60
40
60
SYS

40
20
0
0 20 40 .4 .6 .8 1 2 3 40 60 80 0 20 40 60 80

YIELD NDVI IRRI LPI SYS

NDVI,LPI,Irri, Sys and Yield

Fig : 5.17 - is shown the correlation of all parameters. In this study, found NDVI
and yield are influence in correlation. NDVI and IRRI, LPI, SYS are not influence in
correlation, those are found scatters in correlation. The correlation of yield and other
parameters are influence in good relationship. The study found irrigation frequency is
more, yield also more. Likewise, Sys and LPI are more higher, the yield is also more.

5.5.5 Single date image yield prediction using high-resolution satellite data

The supervised classification images of IRS P-6 LISS – III (3rd March,05) was
used to mask wheat and the NDVI for wheat was generated. The harvesting sites
identified during CCE were overlaid on the wheat mask image and an average NDVI of 3
x 3 pixel window was extracted. Empirical relationships were developed between the
yield obtained from the CCE and corresponding average NDVI to develop single date
yield prediction model for wheat. This study has demonstrated that, with a single date
image, field level yield can be predicted up to( 60.84%). The results of this study shows
at tables 5.4 and 5.5, NDVI (n= 44) from LISS-III was statistically significant with high
coefficient of determination (0.532), low standard error (7.142). In RS and land factors,
Table : showed LPI (n=18) was found minimum 40.4 (Fairly good, class III) and
maximum95.0(Excellent, Class- I).The correlation of LPI and wheat yield was high
significant in determination(0.609).Sys (n=18) was high correlated with wheat yield and

42
its correlation (r2) was 0.661. Likewise, NDVI and LPI (n= 18)was significant high
coefficient of determination( 0.596) low standard error (7.049) and NDVI and Sys
(n=18)was high correlated in determination (0.561) low standard error (7.340). In RS,
Land and management factor input, combination of NDVI, LPI and irrigation
frequency(n=18)was had high coefficient of determination(0.722) low standard error
(6.049). While NDVI, Sys and Irrigation Frequency model had experienced relativity less
variability of wheat yield with a coefficient of determination (0.653).

The result in this study was found using RS, land and management factors model
is better than NDVI alone model in wheat crop yield estimation. The correlation of NDVI
alone and yield was coefficient of determination (0.532).NDVI, land and management
factors were coefficient of determination(more than 0.532).

The study found single date images can provide useful information of the crops
and yield status. But the timing of the image to be used for yield estimation is important.
Though Gielen et al. (2001) explained that there is good correlation between NDVI and
yield but using NDVI as an end-of season yield estimator gives unsatisfactory results
because of the problems of choosing the best time of the image to use, Muthy et al.
(1994) found that vegetation indices calculated from images taken at panicle initiation
and heading stages have high correlation with yield, therefore, they can best be used for
yield prediction.

However, it is difficult to have a single date image representing one phenological


stage at field level because of the differences in planting dates and the varieties used,
resulting in wide differences in crop phenological stages. In this study, provides good
information to predict end-of-season yield as long as it is within the time when there is
maximum vegetation (between panicle initiation and heading stage) and other production
factors are taken into account.

Table – 5.5 Descriptive Statistics of Casual Variables & Its Correlation With Wheat Yield

43
Pearson
Variable Count Min Max Mean SD Correlation

Remote Sensing
NDVI 44 0.44 0.88 0.719 0.118 0.729**

Land factors

LPI 18 40.4 95.0 76.7 19.9 0.609**


SYS Index 18 15.0 79.0 61.0 17.19 0.661**

Management input
Urea applied
14 180 360 276 74.5 0.446
(Kg/ha)
Irrigation
28 1 4 2.57 0.87 0.176**
frequency

** Significant Et 0.01% level (2 failed)

Table 5. 6 : Models developed using RS, land and management factors

Adj.
Variables Count Model fit R2 SEE p
R2
Remote Sensing

NDVI 44 -9.895+60.84*NDVI 0.532 0.521 7.142 0.001

NDVI 18 -1.715+49.0*NDVI 0.432 0.403 8.046 0.003

RS& Land factors


0.001
NDVI,L P I 18 -10.689+37.604*NDVI+0.222*LPI 0.596 0.542 7.049

-4.925+31.495*NDVI+0.256*SYS
NDVI, SYS 18 0.561 0.502 7.34 0.002
Index

RS, Land Factors& Management Inputs

NDVI, L P I,Irrigation 18 -6.148+12.981*NDVI+0.197* 0.722 0.663 6.049 0.00

44
frequency LPI+5.694*Irri
NDVI, SYS, Irrigation -0.787+13.534*NDVI-0.178*SYS
18 0.653 0.579 6.758 0.002
Input index+5.116*Irrigation

( b)
(a)

(c) (d)

Fig-5.18 (a) Land use/land cover map, (b) Wheat mask map

(c) Wheat mask –NDVI map , (d) Yield map

45
Fig – 5.19 Correlation of observed yield and Farmer’s expected yield

The result in this study, predicted yield and observed yield were also high
correlated. In Fig -19: showed independent CCE site (n=2) and farmers expected yield
were high significant in correlation. In this method, CCE sites and farmers expected yield
was statistically significant with high coefficient of determination (5.33 Q/ha).

5.6 Discussion

Through integration of remotely sensed data, land and management factors, an


assessment of field level yield prediction has been executed. It is shown that a
combination of NDVI, land and management parameters can improve field level yield
prediction above use of NDVI alone. It is found that the availability of irrigation
frequency, the number of fertilizer applications, soil type and the date of transplanting
explain yield variability. The date of transplanting and the soil type are found to relate to
the NDVI variability at field level.

The NDVI is the best parameter for crop yield prediction. But if only NDVIs are
used for field level yield prediction, a lot of additional factors explaining yield variability
are not covered. NDVI explains the yield variability caused by transplanting time, soil
type. It does not reflect other production factors like availability of water and other
management factors.

46
The use of land and management parameters alone has shown that the yield
variability can be explained. The combination of NDVI, land and management factors
together are improved the model. This shows that use of NDVI alone, as done
in many studies, can be improved if land and management factors are
also considered, especially at field level where parameters vary from
field to field, as opposed to regional or national level, where these
factors are generalised.

All these findings indicate that there is correlation between remotely sensed
spectral data and yield. The differences in the correlations and explaining ability of yield
variability is due to the level of application and the quality of data being used to
investigate the relationships and to derive models. Muthy et al. (1994) used yield
estimates from CCE, which are fairly accurate and used time composite NDVI. Mohd et
al. (1994) used yield from highly controlled research plots. This study used data collected
through interviews. From the results of this study, it is a significant effect on the degree of
the relationships between remotely sensed NDVI and yield.

5.7 Recommendation

This study has investigated the crop discrimination and field level yield
estimation. The data used in this study was provide by farmers through interview,crop
cutting experiments. The crop yield model in this study was not completed that it was
based on NDVI, some land and management factors and due to project period. There is
need for further research to find more factors that can explain yield variability at field
level and improved model.

There is need for further investigation on the factors that contribute to yield
variability at field level. Factors such as soil type in this study were used from small scale
map. As such, it could not provide detailed variation that may occur at field level. The
study did not consider the previous land management practices, which may have a
residual effect on crop growth and production. There is need for detailed study on the soil
characteristics at field level and include other management factors for specific areas /
fields.

47
6. Conclusion

The Saharanpur district of Uttar Pradesh ha been taken as the study area for this
project with an objective of Crop Yield Estimation of Wheat by Integrating RS, Land and
Management Factors. The result of this study shown that wheat crop is highly separable
and can be discriminated with more than 95% accuracy using high resolution multi-
spectral LISS-III on board IRS-P6 satellite data. A strong linear and non-linear empirical
relation of NDVI and land, management factors has shown possibility of using satellite
NDVI for retrieving yield model for regional productivity analysis.

High-resolution multi-spectral LISS-III satellite data is good for discriminating


crop in the study area. The combination of satellite NDVI, land and management factors
can approve field level yield prediction. Soil sub_ group and NDVI relationship can
provide reasonable accurate yield prediction.

Management factors such as: irrigation frequency and urea apply also can provide
to accurate yield prediction Single date image which is taken at panicle initiation and
heading stages is also provide good information for yield prediction.

48
REFERENCES

Agrawal,R and Jain,R,C., (1982) composite model for forecasting rice yield. Indian
Journal, Agril. Sci., 52 : 177-181

Dadhwal,V.K., Ruhal,D.S., Medhavy,T.T., Jarwal,S.D., Khera, A. P., Singh, J., Sharma,


T., and Parihar, J. S., (1991). Wheat acreage estimation for Haryana using
satellite data. Journal of Indian Society of Remote Sensing, 9 : 295-301.

Dubey, R. P., Ajwani, N., Kalubarme, M. H., Sridhar, V. N., Navalgund, R. R., Mahey,
R. K., sindhu, s. S., Jhorar, O. P., Chrma, S. S., and Narang, R. S., (1994).Pre-
harvest wheat yield and production estimation for Punjub, India, Journal of
Indian Society of Remote Sensing, 15 (10) : 2137-2144

Kalubarme, M. H., Vyas, S. P., Manjunath, K.R., Bhagia, N., Sharma, r., Zadoo, s.,
Gupta, P.C., and Prasad, D.V.V. (1992). Pre-harvest wheat production forecast
for rabi 1990 and 1991-92 ub western Uttar Pradesh using IRS-LISS-I digital
data. In Proceedings of the National Symposium of Remote Sensing for
Sustainable Development, 17-19 Nov. 1992, pp : 349-353

Kumar, S., N.R.Patel., (2004). GIS aided Land Evaluation based on Biophysical
Characteristics using Temporal Satellite Data.

Mahey, R.K., Singh, R., Sidhu,S.S., Narang, r.S., Dadhwal, V.K., Parihar, J.S., and
Sharma (1991), Pre-harvest state level wheat acreage estimation using IRS-IA-
LISS-I data in Punjub. International Journal of Remote Sensing, 14, 1099-
1106

Manhey, R.K.,Sindhu, s.s., Jhorar, O. P.,Chrma, S. S., and Narang, R. S.,(1994). Pre-
harvest wheat yield and production estimation for Punjub, India, Journal of
Indian Society of Remote Sensing , 15(10) : 2137-2144

Manjunath, M.N., Shukla, M., Patel, N.R.,and Pande, L.M., 2002. Discrimination and
Empirical Modeling of Wheat and Sugarcane Crops Using Remote Sensing and
Ground Observations.

Navalgund, R. R., Parihar., J. S., Ajai and Rao P.P.N., (1991). Crop inventory using
remotely sensed data. Current science, 61 : 162-171

Patel, N.K., Singh, T.P., Sahai, B., and Patel, M.S., (1985). Spectral response of rice crop
and its relation to yield attributes. International Journal of Remote Sensing,
6 : 655-664

Pearson, R.L., and Miller, L.D., (1972). Remote Mapping of Standing crop biomass for
estimation of the productivity of the short grass prairie. Pawnee National

49
Grasslands, Colorado. Eighth International symposium on remote Sensing of
Environment, ( Ann Arbor : University of Michigan), 1357-1381

Saha, S.K., and Jonna, s., 1994. Paddy acreage and yield dstimation and irrigated crop
land inventory using satellite and agro-meteorological data, Asian-Pacific
Remote sensing Journal, 6(2), 79-87

Sawasawa, Haig, L.A. 2003.Crop Yield Estimation : Integrating Remote Sensing, GIS,
Management and Land Factors.

Toroev Ddurusbek Isanovich, and Patel, N. R., 2001. Crop Inventory and Soil Suitability
Assessment for Land Use Planning – A Remote Sensing and GIS Approach.

50
Field Data Collection Photos

51

You might also like