You are on page 1of 6

Unsupervised Image Segmentation on Hyperspectral

images for predicting yield and drought in Dhule


region of Maharashtra.
Shreyas Sarasan, Agastya Sanjai, Hemant Palivela, Preeja Babu
Student, Student, Researcn Scholar, Research Scholar.
Mukesh Patel School of Technology and Management Engineering,
Narsee Monjee Institute of Management Studies,
Mumbai, Maharashtra, India- 400057

ABSTRACT- Farmers in India have small land


holdings and due to this, analyzing hyperspectral
images becomes an issue. Due to a high probability
of obstacles in small land holding areas, hyper
spectral images will give less accuracy. So, the
major concern will be to remove obstacles in the
small land holdings by using an unsupervised
segmentation method. The base data set used
consists of hyperspectral images procured from the
earth-explorer website. In the experiment, an
image was first segmented, and its individual
segments are plotted as vertices on a segmentation
graph; which coupled with a corresponding vertex
gives a walk-based graph kernel. The next step in
the process is the Support Vector Machine (SVM) ,
which computes the Normalized Deviation
Vegetation Index (NDVI); which is then used to
compute the Standard Precipitation Indices (SPI).
Now the SPI threshold is applied to understand the
drought severity in the area and NDVI helps in
analyzing the crop yield.
Keywords High-level feature extraction, graph
kernel, walk kernel, support vector machine, NDVI,
SPI.
1. Introduction
For experimental purpose the paper uses hyperspectral
image of a place named Fagne, Dhule in Maharastra
Region of India. The main objective is to compute SPI
values for the Dhule region of Maharashtra, so that the
vegetative yield and drought can be predicted to a high
degree.
All operations are done on hyper-spectral images that
have been obtained from a satellite. The basic strategy
is to homogenously segment each image and compute
segmentation graphs for each. The vertices in the

segmentation graph represent features of a particular


segment, while the edges are used to connect adjacent
segments. A graph kernel is obtained for the
segmentation graphs of the two images to be compared,
and the results are processed through an SVM, which
computes NDVI values. These values are then used in
an equation that outputs SPI. SPI has proven in the past
to be the most reliable measure of drought and land
moisture, and was thus, chosen to be the output index
in this experiment. [7]
2. Method
The experiment is carried out in five phases, each of
which are listed below in detail.
2.1HYPERSPECTRAL IMAGE.
Like all spectral imaging methods, the aim of
hyperspectral imaging is to gain a spectrum collated
from across many bands (including the visible blue,
green and red) for every pixel in an image or a scene.
The general purpose of hyperspectral imaging is to
identify materials, detect objects and find processes. A
hyperspectral image consists of data measured across a
wide spectrum of bands leading to an image that
presents much more data than a normal one does. Many
objects have characteristic spectral emissions (much like
fingerprints), which are used to identify them easily.
Hyperspectral images find application in a wide variety
of fields such as agriculture, physics, mineralogy etc.
Hyperspectral images from the base data set upon
which computations are to be performed. The base data
set was obtained by availing the services offered by the
earth-explorer website [1]. By entering the latitude and
longitude of the desired place in the websites user
interface, one is able to access a vast database of
hyperspectral images captured from various satellites
stored in different data-sets, with these images being
obtained from the EROS Moderate Resolution Imaging

Spectro-radiometer (eMODIS) data set for Dhule,


Maharashtra. (Latitude- 20.9042, Longitude- 74.7749)

2.2 OVERVIEW OF THE PROJECT. (LITERATURE SURVEY


INCLUDES SVM AND JSEG)

A general overview of the entire process can be


represented in a series of steps as given below:
Images from the data set are homogeneously
segmented
(unsupervised),
and
segmentation graphs for each of them are
computed.
The second stage is kernel computation [1]
[2], where the first step is computing a kernel
for two corresponding vertices in the
segmentation graphs. This is succeeded by
kernel computation for specific walks within
the segmentation graphs, thereby obtaining a
walk-based graph kernel. The next step is the
summation of walk-based kernel for walks
of varying lengths, which in the end results
in a positive definite kernel.
This kernel is then processed by the SVM
and classification is done. [2]
The final stage involves the calculation of
NDVI [4] values and the subsequent
calculation of SPI by the use of the relation
that exists between NDVI and SPI.
The main focus of this paper is on Stage 1, where
image segmentation takes place; while the rest of the
stages shall be explained in future works.

The applications of image segmentation as a method


are many, and the ways in which it can be achieved
continues to be a subject of increasing endeavors. Image
segmentation in the scope of this work has been
achieved by the use of an algorithm known as JSEG [2]
(Explained in detail in Section III.).
Before the segmentation process starts, a few
assumptions are made about the images to be processed:
Each image contains a set of approximately
homogeneous color-texture regions.
Color information in each image region be
represented by a few quantized colors.
Colors between two neighboring regions are
distinguishable.
Since it is rather resource intensive and difficult to
analyze color similarity and distribution at the same
time, the entire process is divided into two parts; the first
being color quantization and the second being spatial
segmentation.
Color quantization [8] is done in color space without
taking into account the spatial distribution of the colors.
After this, all pixel values are changed to their
corresponding color class values, thereby resulting in a
color class map, which resembles a specialized texture
composition map. Color quantization works on the basic
principle that human vision is more perceptive to
changes in smooth areas than it is in rough (textured)
areas. This lends itself to the realization that colors in
rugged (textured) areas can be more coarsely classified;
this being achieved by perceptively weighing individual
pixels. This is achieved as follows:
With the aim of smoothening the image, peer
group filtering is applied. [10] (A non-linear
algorithm.)
This produces values that indicate the
smoothness of a local area.
Values are assigned to each pixel based on
whether its located in a smooth area or a
textured one. The ones in the textured area
are weighted less than the ones in the smooth
area.
CIE LUV color space is used for its
perceptual uniformity, and color vector
quantization is achieved through a modified
version of General Lloyds algorithm.
The next step is spatial segmentation [8] and this is
done without regard to corresponding pixel color
similarity, which is succeeded by the computation of J
images, which are approximations of local

homogeneities at different scales, which can indicate


boundary locations.
When several homogeneous color regions exist,
there is more separation between color classes, and the
value of J (forming the base unit of the J-image) is large.
In case of color regions distributed more uniformly, the
value of J is small. When J is calculated for each region
instead of the entire class map, the average of all J values
is given by . The value of is inversely related to better
segmentation. Lesser the value, better the segmentation.
The next stage in the entire process is region
growing, where some local values of J (minima) are
selected arbitrarily and act as seeds. Before choosing
the seeds, meeting some criterions ensures good
selection of seeds. These criterions are presented below:

Calculation of average and standard


deviation of local J values in a local region.
Calculating a threshold defined as the sum of
standard deviation and a*average deviation,
a chosen from several presets.
If local j values of a pixel are found to be
under the threshold, it is a potential seed
candidate.
The regions are then grown from the seeds, each
region growing pixel by pixel. Region growing often
leads to over-segmentation, this being the result of local
area color similarity. To remedy this, an agglomerative
method is used [9] that merges two histograms with the
least distance between them. This continues to be done
until a maximum distance threshold is achieved.

Figure 1: JSEG Algorithm

J-images correspond to measurement of local


homogeneities at different scales, which can indicate
potential boundary locations. The newly constructed
For the case of an image consisting of several
homogeneous color regions, the color classes are
more separated from each other and the value of
J is large. , if all color classes are uniformly
distributed over the entire image, the value of J
tends to be small. The motivation for the definition
of J comes from the Fisher's multiclass linear
discriminant [10].

2.3THE JSEG ALGORITHM.

JSEG [8] is presented for the fully unsupervised


segmentation of color-texture regions in images
and video. The segmentation consists of color
quantization and spatial segmentation. A criterion
for good segmentation is proposed. Applying the
criterion to local image windows results in J-images,
which can be segmented using a multi-scale region
growing method.

Figure 2: Region Growing Expansion

The characteristics of the J-images allow us to use a


region-growing method to segment the image.
Consider the original image as one initial region.
The algorithm starts the segmentation of the image
at a coarse initial scale. Then, it repeats the same
process on the newly segmented regions at the next
finer scale.
Region growing consists of determining the seed
points and growing from those seed locations.
Region growing is followed by a region merging
operation to give the final segmented image.
Final output depends on three parameters given by the
user. Threshold for color quantization, number of scales
for each image, threshold for region merging.The
image thus processed gives us a Segmented
Hyperspectral image where the regions are
separated.
3. Literature Survey
Let Z be the set of all N data points in a class-map.
Let = (, ), and m be the mean [8],
1
=
. . (1)

Suppose Z is classified into Z classes,


, = 1, ,C.
Let be the mean of the data points of class
,
1
= . . . (2)
N

4. Existing System
The current model of Standard precipitation index
in the Dhule region/India, is time consuming and
not useful for short term prediction in India.
Extensive field work is carried, the atmosphere is
studied along with a determination of conducive
elements present in the soil by testing it.
In countries like Germany or Iran, Huge Land
Holdings are present, where manmade structures
are not allowed to be built. The land is vast and
natural structures and ecosystems of an organic
nature, are the preexisting elements. There is no
interference from the side of man.
A study done on drought conditions in Iran, for
example, using Images obtained from Satellites
and the use of filters and parameters such as
NDVI, NDVI-SEG, VCI, TCI [11] were used to
find the output, SPI. During the training of the
neural network, it would not have to face problems
regarding man made structures which could
reduce the accuracy or prove a setback to the
objective to predict drought. Therefore, it is very
easy to use Multilevel Perceptron (Best neural
network for such countries) or Radial Based
Functions (RBFs) and even SVM classification.
These man-made structures can interfere with our
objective to segment and image and draw
meaningful conclusions from them.
5. Proposed System
existing system)

(Advantages

over

Let
= || ||2 . . (3)

And,
=

= || || . (4)
=1

is the total variance of points belonging to the


same class. Define

=
. . . (5)

We will now recalculate J over each segmented


region instead of the entire class map and define
the average by,
1
=
. . (6)

Using Images obtained from earth-explorer, we


apply JSEG algorithm to obtain homogenous color
texture regions and we thereby obtain Segmented
Images which are classified by the SVM by Kernel
Calculation.
Reason being, this classification helps in high level
feature abstraction by first identifying and the manmade objects like electric poles, Man-made
structures, roads and even natural structures such
as landscapes, grass, forests, sandy areas, low lying
areas etc., which help us focus on extracting the
relevant conditions which are precursor to extract
NDVI parameters/indices and obtain a SPI as our
output to predict drought condition/drought.

5.1 Results.

Figure 3: Original Hyperspectral


Image of Dhule area

Figure 4: Segmented image of phase 1

Figure 4: Shows the output of region


growing algorithm of phase2

Figure 5: Shows the output of region


growing algorithm of phase3

Figure 3 shows a clear hyperspectral image that has been


procured from earth explorer. Paper applies phase 1 JSEG
algorithm to the existing image so as to get the unsupervised
cluster assignments.

Figure 6: shows the output of region growing


algorithm of phase4

Figure 3 gives the look over of how the region growing


algorithm works with a scale rate of 0.25. Once the
threshold is addressed the region merging takes place.
In the next phase of the project the segments will be
assigned to each vertex of the segmentation graph. Each
vertex will be assigned set of variables that form a
column vector. Each and every column vector is
designed with 23 different features like x, y, intensity,
saturation etc.
6. Conclusion
Images were segmented using the process applied with
respect to the JSEG algorithm to obtain homogenous
color-texture regions. In future work, the concept of
walk based graph kernels will be used to calculate
multiple images, leading to extraction of NDVI indices
from the hyperspectral images. These NDVI values will
be then used to calculate the SPI values to give a better
perspective on the vegetative yield and drought
probability.
7. REFERENCES

1. Harchaoui, Zad, and Francis Bach. "Image


classification with segmentation graph
kernels." Computer Vision and Pattern
Recognition,
2007.
CVPR'07.
IEEE
Conference on. IEEE, 2007.
2. Vert, J-P., et al. "High-level feature extraction
using SVM with walk-based graph kernel."
Acoustics, Speech and Signal Processing,
2009. ICASSP 2009. IEEE International
Conference on. IEEE, 2009.
3. Gautam, R. S., et al. "Application of SVM on
satellite images to detect hotspots in Jharia coal
field region of India." Advances in space
research 41.11 (2008): 1784-1792.
4. Vuolo, Francesco, and Clement Atzberger.
"Application of MODIS NDVI time series data
for land cover classification at European scale."
EGU General Assembly Conference Abstracts.
Vol. 15. 2013.
5. Jain, Sanjay K., et al. "Application of
meteorological and vegetation indices for
evaluation of drought impact: a case study for
Rajasthan, India." Natural hazards 54.3
(2010): 643-656.

6. Marshall, Geoffrey, and Xiaobing Zhou.


"Drought detection in semi-arid regions using
remote sensing of vegetation indices and
drought indices." Geoscience and Remote
Sensing Symposium, 2004. IGARSS'04.
Proceedings. 2004 IEEE International. Vol. 3.
IEEE, 2004.
7. Al-Qinna, Mohammed I., et al. "Drought
analysis in Jordan under current and future
climates." Climatic change 106.3 (2011): 421440.
8. Yining Deng, B.S Manjunath. Unsupervised
Segmentation of Color-Texture Regions in
Images and Video on IEEE.
9. R.O. Duda and P.E. Hart, Pattern Classification
and Scene Analysis,New York: John Wiley &
Sons, 1970.
10. Y. Deng, C.Kenney,
M.S. Moore,
and
B.S.Manjunath, Peer Group Filtering and
Perceptual Color Image Quantization, Proc.
IEEE Int'l Symp. Circuits and Systems, vol. 4, pp.
21-24, 1999.
11. Mahdi Jalili, Joobin Gharibshah, Seyed Morsal
Ghavami, Mohammadreza Beheshtifar, Reza
Farshi, Nationwide Prediction of Drought
Conditions in Iran Based on Remote
Sensing
Data. On IEEE

You might also like