GPU ACCELERATED AUTOMATED FEATURE EXTRACTION FROM SATELLITE IMAGES

International Journal of Distributed and Parallel Systems (IJDPS) Vol.4, No.2, March 2013
DOI : 10.5121/ijdps.2013.4201 1
GPU ACCELERATED AUTOMATED
FEATURE EXTRACTION FROM SATELLITE
IMAGES
K. Phani Tejaswi, D. Shanmukha Rao, Thara Nair, A. V. V. Prasad
National Remote Sensing Centre, Indian Space Research Organization, Hyderabad, India
tejaswi_phani@nrsc.gov.in,thara_nair@nrsc.gov.in
ABSTRACT
The availability of large volumes of remote sensing data insists on higher degree of automation in feature
extraction, making it a need of the hour. Fusing data from multiple sources, such as panchromatic, hyper
spectral and LiDAR sensors, enhances the probability of identifying and extracting features such as
buildings, vegetation or bodies of water by using a combination of spectral and elevation characteristics.
Utilizing the aforementioned features in remote sensing is impracticable in the absence of automation.
While efforts are underway to reduce human intervention in data processing, this attempt alone may not
suffice. The huge quantum of data that needs to be processed entails accelerated processing to be enabled.
GPUs, which were originally designed to provide efficient visualization, are being massively employed for
computation intensive parallel processing environments. Image processing in general and hence automated
feature extraction, is highly computation intensive, where performance improvements have a direct impact
on societal needs. In this context, an algorithm has been formulated for automated feature extraction from
a panchromatic or multispectral image based on image processing techniques. Two Laplacian of Guassian
(LoG) masks were applied on the image individually followed by detection of zero crossing points and
extracting the pixels based on their standard deviation with the surrounding pixels. The two extracted
images with different LoG masks were combined together which resulted in an image with the extracted
features and edges. Finally the user is at liberty to apply the image smoothing step depending on the noise
content in the extracted image. The image is passed through a hybrid median filter to remove the salt and
pepper noise from the image. This paper discusses the aforesaid algorithm for automated feature
extraction, necessity of deployment of GPUs for the same; system-level challenges and quantifies the
benefits of integrating GPUs in such environment. The results demonstrate that substantial enhancement in
performance margin can be achieved with the best utilization of GPU resources and an efficient
parallelization strategy. Performance results in comparison with the conventional computing scenario have
provided a speedup of 20x, on realization of this parallelizing strategy.
KEYWORDS
Graphics Processing Unit (GPU), Laplacian of Gaussian (LoG), panchromatic.
1. INTRODUCTION
The massive increase in the dimension of remotely sensed data that is being obtained by an ever-
growing number of earth observation sensors, has led to a condition, where there is large number
of images to be processed. To tide over this scenario, automated feature extraction techniques can
be employed for the extraction of geo-spatial features. Automated feature extraction can be defined
as the identification of geographic features and their outlines in remote-sensing imagery through
post-processing technology that improves feature definition, either by increasing feature-to-

2
background contrast or by the usage of pattern recognition software [1]. This is an important task
in numerous applications ranging from security systems to natural resource inventory based on
remote-sensing. Image interpretation is an important and highly challenging problem with
numerous practical applications.
Edge detection is one of the most commonly used operations in image analysis and there are many
algorithms for enhancing and detecting edges .Edges form the boundary between an object and the
background, and indicates the boundary between overlapping objects. Traditional approaches use
classical edge detectors viz. Sobel operator, Canny Edge detector, Prewitt edge detector and LoG
operators that work fine with high quality pictures. Of these the LoG operator is the most
commonly used. Different features of interest can be extracted by varying the LoG filter size.
Filter of high variance yields the finer edges of the image where as filter of low variance gives the
inner details of the image.
Zero crossings lie on closed contours. Zero crossing detectors provide a binary image with single
pixel thickness lines indicating zero crossing points. 'Edges' in the images appear as zero crossings
and can be detected by sign change in LoG in at least one direction. When the contrast with respect
to the neighbors of the pixels is very high, a zero crossing will be detected and can be extracted to
identify and demarcate the edges more clearly. The standard median filter [2] is a simple rank
selection filter that removes impulse noise. This is achieved by replacing the luminance value of
the center pixel of the filtering window with the median of the luminance values of the pixels
contained within the window. The drawback of median filters is that, it removes thin lines and
blurs image details at low noise densities. Median filtering technique is generally employed to
eliminate the salt and pepper noise from an image. The hybrid median filter is a modified median
filter that draws the luminance values from the pixels parallel to and at 45deg to central pixel of
the filtering window. This has been proposed to avoid the inherent drawbacks of the standard
median filter by controlling the tradeoff between the noise suppression and detail preservation [3].
This paper discusses the edge detection, focusing on LoG, zero crossing detection and various
statistical techniques to demarcate the edges and to extract the different features in panchromatic
data, applicable for urban area detection. Combining these image processing approaches in a
panchromatic image or a single band in a multispectral environment enables us to extract features
of interest viz. water bodies, roads, buildings etc. The related work carried out earlier in this field
is discussed in second section and the theoretical reasoning behind the selection of LoG operator,
zero crossing threshold, removal of salt and pepper noise and their applicability in this
development are explained in the succeeding section. Fourth section describes the automated
feature extraction algorithm and its implementation details. Subsequent sections give an insight on
the General Purpose Graphical Processing Units (GPGPUs) and their effective exploitation by
adopting various design approaches. Detailed evaluation of the performance of the finalized
optimal approach is discussed with the performance margin and the results on various test images.
2. RELATED WORK
Feature extraction of satellite images is of major interest in a slew of applications as it is the major
input for the remote sensing applications. Edge detection employing techniques is also dominant
in various image feature extraction techniques in which 2-D wavelet transforms are applied on the
image for several levels [4]. Several existing algorithms on edge detection like Marr-Hildreth edge
detector, Canny edge detector [5], make use of Laplacian of Gaussian operator with a single
exclusive sigma value. Canny edge detector has been efficiently used in the extraction of features
from the satellite images [6]. It allows a pre-processing step of Gaussian smoothing prior to actual
application of LoG filter. Median filtering [7] is the most commonly used approach for the

3
removal of salt and pepper noise [8] from the image. The ranking of the neighboring pixels
depends on the brightness and the median value is now made the new value of the central pixel in
a hybrid median filter. Most of the edge detection techniques and feature extraction algorithms
were designed for IKONOS images [9,10] and similar with a very high resolution of the order of 1
m.
The approach discussed in this paper targets Cartosat-1 and Resourcesat AWiFs images which are
lower resolution images. The same algorithm can be applied for extracting the features of the
satellite images of very high resolutions. The LoG operator followed by zero crossing detection
has been used as the edge detection technique where no significant details of the image are lost
compared to the wavelet approach. The drawback in the application of wavelet technique is that
the output resolution decreases with increase in wavelet levels.
Another unique feature of this algorithm is that it employs the LoG operator with two different
sigma values, a high sigma to extract the features of interest and that of low sigma to extract the
edges in the image.
Yet another highlight is the selection of hybrid median filter for noise removal. The disadvantage
of the conventional median filtering techniques is they tend to erase lines narrower than half the
width of the neighborhood. Hybrid median filtering is chosen as the technique for the removal of
salt and pepper noise in the image as it gets along the limitations of conventional median filtering
technique and involves a three-step ranking process that uses two subgroups.
This is totally unsupervised and can be applied with ease to any satellite image of any resolution
unlike many algorithms which employ supervised classification in which the pixels need to be
trained a priori.
3. SCIENTIFIC RATIONALE
3.1. LOG
The Laplacian of an image highlights regions of rapid intensity change and is therefore often used
for edge detection. These kernels apply a second derivative measurement on the image and hence
are highly susceptible to noise frequencies. Using any standard kernel, the Laplacian can be
calculated using standard convolution methods. To counter this, the image is often Gaussian
smoothed before applying the Laplacian filter. Smoothing by Gaussian helps to reduce the high
frequency noise components ahead of differentiation. The LoG operator detects and extracts the
edges from an image. The suitable mask can be selected depending on the requirement to extract
the edges or the features of interest. The 2-D LoG function centered on zero and with Gaussian
standard deviation σ has the form:
(1)
By itself, the effect of the filter is to highlight edges in an image. σ also controls the size of the
Gaussian filter. A higher value for σ results in a larger size of Gaussian filters. This implies
detecting larger edges. Smaller values of σ imply a smaller Gaussian filter maintains finer edges
in the image. The selection of variance and window size can be used to provide edges at various
scales. LoG operator omits low and high pass frequencies, it is equivalent to band-pass filter.

4
Choice of the value of variance controls the spread of operator in the spatial domain, setting
variance to a high value gives low-pass filtering, as expected.
The advantage of employing LoG operator in edge detection is that it helps in finding the correct
places of edges and testing wider area around the pixel. The user can tailor the algorithm by
adjusting the variance to adapt to different environments.
3.2. Zero crossing
An approach to finding edges is to detect zero-crossings in second derivative or LoG filtered
output. The zero crossing is computed by observing the four neighbors of each pixel. If they all
have the same sign as candidate pixel, then no zero crossing is detected. On the contrary,
candidates having the smallest absolute value compared to its neighbors with opposite sign are
indicative of zero crossings. Subsequently the zero-crossings are to be passed through a threshold
to retain only those with large difference between the positive maximum and the negative
minimum, so as to suppress the weak zero-crossings most likely caused by noise.
3.3. Standard Deviation
Sample standard deviation is defined by the unbiased estimate of the standard deviation, sa, of the
brightness within a region (ℜ) with Λ pixels and is denoted by:
Standard deviation of a pixel in an image in general can be computed as:
= (2)
where ma is the mean of all values in the data set.
The standard deviation plays an important role in identifying the edges of the image
which has been threshold with a zero crossing detector.
3.4. Median Filtering
Median filters are non-linear digital filters known for their capability to remove impulse noise,
with minimum signal distortion and without hampering the edges. In particular, compared to the
smoothing filters examined thus far, median filters have many advantages viz. no reduction in
contrast across steps, no shifting of boundaries and no occurrence of new unrealistic values as
new values are not created near edges. The degradation of edges is minimal in median filtering,
which makes it possible to apply median filters repeatedly. A specialized median filter is the
Hybrid Median filer. This adopts a 3-step ranking process using two subgroups of a 5x5
neighborhood. These subgroups are extracted from the pixels parallel and at 45o
to the edges,
with reference to the center reference pixel. The median for each subgroup is determined.
Comparison is carried out between the two values and the original pixel value and the median of
these three values forms the output value for the pixel in the selected window in the filtered image. The
window size can be selected as required by an application.
4. ALGORITHM FOR AUTOMATED FEATURE EXTRACTION
An algorithm which employs image processing techniques has been formulated which enables us

5
to extract information like roads, buildings or water bodies from any panchromatic image or any
single band of a multispectral image.
The Laplacian of Gaussian filter is used to extract the edges. Depending on the value of variance,
the size of the Gaussian filter varies. A larger Gaussian filter detects larger edges and a smaller
variance can be employed to detect finer edges in the image. Hence incorporating a proper
combination of these two filters helps us to detect all edges in an image.
Applying a LoG filter alone will not provide edges with the required clarity. Once the edges are
filtered through the LoG filter, a zero crossing threshold needs to be applied on the LoG filtered
image. The zero crossing detector with a predefined threshold scans through the four neighbors of
each pixel. Pixels with the smallest absolute value compared to its neighbors with opposite sign are
qualified as zero crossings, once they are cleared through a threshold detector. This stage defines
the larger and finer edges of the image, which is to be followed by identifying and marking the
required features explicitly.
Identifying and segregating the region of interest was the succeeding requirement. Since the zero
crossing threshold detectors was applied on the output of LoG filter, statistical techniques like
standard deviation had to be applied on matrices with specified dimensions to isolate the larger and
finer edges. The concept of variation in standard deviation values was used to isolate the presence
or absence of edges in a 5x5 pixel neighborhood.
The statistical output might have a large presence of salt and pepper noise, which has to be filtered
properly. Median filter is the best non-linear digital filter since they remove impulse noise without
damaging the edges. Repeated invoking of hybrid median filter with varying dimension and a three
stage ranking process yields an image with trivial noise content.
4.1. Implementation Details of Urban Area Detection Algorithm
The input image was padded with 2 rows/columns on all the boundaries so that each pixel in the
original image had its corresponding 5x5 neighborhood. The Laplacian of Gaussian filters of
variance 0.5 and 20 were applied on the image separately to extract the features and edges
respectively. The LoG filter dimension selected in both the cases was 5x5. The LoG mask was
applied on each pixel with its 5x5 neighborhood and the LoG output images were obtained for the
two filters. A zero crossing detector with a preset threshold value was applied on both the LoG
output images which were padded with 1 row/column on all the boundaries. The zero crossing
logic was applied on each pixel and its corresponding up, down, left and right neighbors. On the
application of the zero crossing detectors, features of interest were demarcated in the low variance
filtered image and finer edge transitions were identified in the high variance filtered image.
Standard deviation value was worked out for each pixel in its 5x5 neighborhood on the zero
crossing output images, padded with 2 rows/columns on the boundaries. If the standard deviation
exceeded the corresponding pre-specified maximum threshold deviations, the standard deviation
was re-calculated for the corresponding pixel with a localized 3x3 neighborhood. Pixels which
cross the set threshold deviation were identified. These pixels denote the features of interest in the
high variance filtered image where as they symbolize the edges in the low variance filtered image.
The features and edges extracted from the low and high variance filtered images respectively were
combined together to achieve the urban area mapped image. The steps involved in urban area
detection are given in the flow chart below

6
Figure 1. Flow Chart for Urban Area Detection
4.2. Implementation Details of Water body ExtractionAlgorithm
The urban area detection algorithm was tailored for the extraction of water bodies in any band of a
multispectral image. The feature extracted image has a characteristic salt and pepper noise which
was removed by passing through a hybrid median filter in multiple levels of higher and lower
dimensions. The steps involved in water body extraction are given in the flow chart below
Start
Padded 2 rows/columns on all boundaries
Zero crossing detector(threshold = 250) was applied on LoG filtered images
Input image filtered through a LoG mask with variances 0.5 and 20
Standard Deviation was computed for a 5x5 neighborhood of every pixel
Input image was copied into a 2-D array in host and copied to GPU global
memory
Padded 1 row/column on all boundaries
Combined the two images to extract edges(std=50) and regions of
interest(std=80)
Image with extracted edges and regions of interest was copied from GPU back
to CPU
End

7
Figure 2. Flow Chart for Urban Area Detection
5. GPU- INTRODUCTION, NEED AND OPTIMIZATION STRATEGIES
It is cumbersome to manage the large volume of remote sensing data to be processed using
automated feature extraction. For larger images, where significant performance improvement may
be required, computation making use of traditional processors may not yield the desired
performance. As a result, deployment of GPUs (Graphics Processing Units) has attained
momentum in applications where large volume of data needs to be processed. In this context, the
application of General Purpose Computing using Graphical Processing Unit (GPGPU), a co-
processor to CPU, for automated feature extraction is being thought of. With remarkable
improvements in GPU computing, due to the improved hardware potential and programmability,
this paper depicts how a GPU, with parallel design architecture can enhanced computational power
significantly to match the requirements of handling enormous data volume[11].
The programming platform that has been selected is NVIDIA Corporation’s CUDA (The Compute
Unified Device Architecture) [12]. The most important advantage of CUDA over most other GPU-
Start
Zero crossing detector (threshold = 250) was applied on
LoG filtered images
Input image filtered through a LoG mask with variances 0.5
and 20
Standard Deviation was computed for a 5x5 neighborhood
of every pixel
Input image was copied into a 2-D array in host and
copied to GPU global memory
Padded 1 row/column on all boundaries
Combined the two images to extract edges (std=1000) and
regions of interest (std=1800)
Image with water bodies extracted was copied from GPU
back to CPU
End
Smoothen the image using hybrid median filter
(successively of sizes 21 and 7)

8
APIs is that it can easily access the memory of GPUs and also permits direct access of the global
and shared memory.
Well-organized implementation of an algorithm in the GPU involves extracting the parallelism of
the target application, and, additionally, adopting efficient data transfer, memory and thread
management techniques. The threads must be scheduled competently and synchronization between
the threads is also essential. Equally important is the efficient use of various memories, viz. texture
memory, constant memory, shared memory, global memory and local registers. All these can
contribute in a significant performance improvement.
A challenging requirement is to parallelize the algorithms to be implemented in GPU. Significant
efforts are needed to refactor the sequential algorithms to achieve the necessary computational
margin in parallel scenario.
The effectual bandwidth of the computations should be considered as a metric when measuring
performance and optimization benefits. To achieve improved performance gain, data transfer
between the CPU and the GPU is to be minimized.
The GPU used for this design set up is nVIDIA Tesla C2075 with 448 CUDA cores [13].
6. DESIGN APPROACH
The test image which is discussed here for urban area detection is a panchromatic image of
Cartosat-1 with 2.5 meter resolution and the corresponding image for water body detection is an
AWiFS image of Resourcesat-2 with 12-bit radiometric resolution and data acquisition in four
spectral bands, B2, B3, B4 and B5 (0.52-0.59; 0.62-0.68; 0.77-0.86; 1.55-1.70 µm) with 56m
spatial resolution. The test images are of sizes, 4000x4000 pixels and 2000x2000 pixels.
The automated feature extraction algorithm had essentially four techniques to be implemented on
the image so as to identify the region of interest. These techniques were implemented on a kernel
mask of 5x5 and extended to the entire image. This modus operandi was extremely time-
consuming and hence it was indispensable to realize alternate efficient parallelization strategy to
accelerate the entire processing chain.
6.1. Design Approaches with CPU
The initial design technique was to employ conventional sequential code in a single core
processor. Highly computation intensive steps discussed earlier will be exceedingly time
consuming, as evident from the succeeding table (Table 1).
The execution speed for all the steps in the automated feature extraction algorithm can be
increased by the exploitation of the inherent parallel resources of a multi-core processor. The
current CPU is an 8-core processor and all its cores can be exploited effectively by dividing any
application or algorithm into modules that are independent of each other and executing each
portion in each core. This can be done using the concept of POSIX or pThreads.
In this algorithm , the steps of LoG filtering, zero crossing detection followed by feature extraction
using standard deviation were duplicated on the raw image, once with the filter of high sigma and
then with the filter of low sigma. Since these were independent of each other, these were executed
in parallel using the multi-core processor. This involved invoking two parallel threads, each
independently executing the steps. The results of both were combined to extract the features and

9
edges resulting in a significant performance improvement as detailed below. The time gain
obtained by the parallel approach employing CPUs is detailed in Table 1.
Table 1. Comparison of approaches CPU
Image Details Conventional CPU approach Parallel approach employing CPUs
Time (ms) Time (ms) % speed-up
Cartosat 1
2000x2000
5950 3420 74%
Cartosat 1
4000x4000
22030 12010 83.5%
AWiFS
2000x2000
42838 24500 74%
6.2 Design Approaches with GPU
The input image available in CPU (host) memory was copied to GPU (device). The entire
processing is carried out in GPU and the detected region of interest is sent back to CPU, after the
processing is completed.
The following sections discuss about the GPU implementation of the Automated Feature
Extraction algorithm employing various optimization approaches and analysis of the throughput
achieved in each case by taking different parameters into consideration.
6.2.1 Row-wise Parallel Approach
An approach that involved assigning a row of pixels to each thread was attempted. In this
approach, an entire row of the image was processed by an individual thread. The total number of
threads was chosen to be the total number of rows, with each thread handling each row of pixels.
The maximum number of threads/block is 1024 and the optimum number is 512, which was
chosen and the number of blocks exercised depended on the image size. The threads were
organized into kernel launch parameters as given in Table 2. This approach has proved faster
compared to the parallel CPU approach and has yielded a considerable throughput that is evident
in Table 3.
Table 2. Kernel Launch Parameters for Row-wise Parallel Approach
Image Size No of threads No of Blocks No of Grids
2000x2000 512 4 1
4000x4000 512 8 1
Table 3. Comparison of Row-wise Parallel GPU Approach with CPU
Image Details CPU Execution Time
(ms)
GPU Execution Time
(ms)
% Speed-up
Cartosat 1
2000x2000
3420 1000 242
Cartosat 1
4000x4000
12010 1400 757.86
AWiFS 24500 1560 1470.51

10
2000x2000
6.2.2 Pixel-wise Parallel Approach
Efficient utilization of GPU resources can be achieved by designing the flow of execution in such
a way that the logic will be implemented on all the pixels of the image at a time, in parallel. The
total number of threads was chosen to be equal to the total number of pixels in the image (number
of rows multiplied by number of columns in the input image), with each thread handling each
pixel.
A new concept of global thread indexing has been employed in this approach which can be
visualized below in Figure 3 (explained with reference to an input image of 2000x2000 pixels).
. . . . . ............... . . .
. . . . . ............... . . .
. .
. .
. .
. .
. .
. .
. .
. .
. .
. . . . . ............... . . .
. . . . . ............... . . .
Figure 3. Global thread indexing
Using the equation (Gthid/2000)*(2000) + (Gthid%2000) for the image of size 2000x2000, all
the pixels in the image were accessed and the logic was executed in parallel for all the pixels.
6.2.3 Selection of Kernel Launch Parameters
The number of threads and the number of blocks is chosen so as to minimize the execution time.
Various threads and block combinations viz. 256/16384, 512/8192 and 1024/4096 were tested
which yield different execution times as shown in the table (Table 4) below
Table 4. Kernel Launch Parameters for Pixel-wise Parallel Approach
Image Size No. of
threads
No. of
blocks
No. of
grids
Execution
Time
(ms)
2000x2000 256 16384 1 457
512 8192 1 446
1024 4096 1 520
Gthid % 2000
(Gthid/2000)
* 2000
If Gthid < (2000*2000)
{
d_buf[(Gthid/2000)*(2000)+(Gthid%2000)]…
.
}
---> Algorithm will be applied for all pixels in the
image in parallel, thereby achieving a very high
throughput.

11
The above table clearly depicts that the execution time was minimum for a block size of 512
threads and hence was finalized.
6.2.4 Optimization Strategies Incorporated
The entire operations involving the edge detection and detecting the regions of interest were done
with a single kernel call from the host. The optimization strategies adopted are detailed in Table 5
Table 5. Optimization Strategies
Optimization Strategies
Adopted
Description
Parallelization Strategy
Involves assigning one pixel per thread, so that the
number of threads is equivalent to the number of
pixels in the input image
Threads per block Number of threads per block was chosen as 512
Multiple Kernels
For improved performance margins, multiple
kernels were invoked within the device, to reduce
the load on each kernel considerably
Block Synchronization
Synchronization between different row-wise and
column-wise operations was carried out using
syncthreads ()
7. PERFORMANCE EVALUATION
The developed GPGPU application was implemented in the following workstation:
CPU : HP Z-800 Intel Xeon 8 core E5620 @2.4Ghz
GPU : nVidia Tesla C2075 with 448 CUDA cores
Subsequent to the analysis of a variety of design approaches in CPU and GPU, we have zeroed
upon the following design configuration which yielded the maximum performance margin.
Invoking multiple kernels with a pixel-wise parallel approach in GPU having 512 threads per
block provided us the best approach. This section provides the results for various test images with
reference to kernel execution timings, data transfer timings and the speed up attained.
The succeeding table (Table 6) provides the data transfer time (host to device and device to host
transfer times combined together) and the kernel execution time for the test images
Table 6. Details of GPU execution timings
Image Details Data Transfer Time
(ms)
Kernel Execution Time
(ms)
Cartosat 1
2000x2000
10 69
Cartosat 1
4000x4000
44 372
AWiFS
4000x4000
6 825

12
This is the most efficient approach compared to its counterparts i.e., conventional CPU approach,
parallel approach using CPU (pThreads) and the GPU approach of on row of pixels per thread. The
efficiencies of various approaches are tabulated in Table 7. The speed up obtained is used as the
measure of throughput.
Table 7. Comparison of Pixel-wise parallel GPU approach with pthreads in CPU
Image Details CPU --- pthreads
Time(ms)
GPU --- Pixel-Wise Parallel
Time(ms)
% Speed-up
Cartosat 1
(2000x2000)
3420 446 666.82
Cartosat 1
(4000x4000)
12010 990 1113.13
AWiFS
(2000x2000)
24500 1182 1972.76
0
500
1000
1500
2000
2500
3000
3500
4000
A B C
CPUParallel
Approach
GPU-Row-wise
Parallel
GPU-Pixel-wise
Parallel
Figure 4. Representation of throughputs of various approaches
The following figures, Figure 5 - Figure 8, denote the Cartosat test image with 2000x2000 pixels
for urban area detection, the features isolated in Figure 6, detected edges in Figure 7 and the final
image showing the output of automated feature extraction which detected roads and buildings, in
Figure 8. Figure 9 and Figure 10 denote the input and output of a segment of 4kx4k image.
Figure 11 - Figure 14, depict the AWiFS image with water bodies, the extracted features and edges
in Figure 12 and Figure 13 respectively and the final extracted water bodies in Figure 14.
Figure 5. Input image for Urban Figure 6. Image with features
isolated Area Detection (2Kx2K)
A : Cartosat 1
2000x2000
B : Cartosat 1
4000x4000
C : AWiFS
2000x2000

13
Figure 7. Image with extracted edges Figure 8. Image with features extracted
Figure 9. Input image for Urban Figure 10. Image with features extracted
Area Detection (4Kx4K)
Figure 11. Input AWiFS image for Figure 12. Image with features isolated
water body Detection(2Kx2K)
Figure 13. Image with extracted Figure 14. Image with extracted water
edges bodies

14
The efficiency of the parallelization strategy adopted can be measured by comparing the GPU
execution time with the CPU execution time. A maximum percentage improvement of 1972% as
compared to parallel approach using multi-core CPU was obtained with this parallelization
methodology.
8. CONCLUSION
The Graphical Processing Units are exceedingly capable parallel processors which acquires data
from the CPU to carry out repetitive calculations on huge volumes at an extremely rapid rate. The
processed data would then be sent back to the host. The discussed GPU design configuration has
provided a noteworthy performance improvement compared to even a multi-core CPU processing.
This developed algorithm has successfully extracted the features and edges from panchromatic as
well as multi-spectral images with relatively large resolutions as well. This can be further extended
to extract any specific feature including roads, buildings, water bodies, snow, cloud etc., from a
panchromatic image or a multispectral image using textures. This will help in isolating and
classifying diverse features with enhanced accuracy.
ACKNOWLEDGMENTS
The authors desire to acknowledge the support and excellent guidance made available by Dr. V. K.
Dadhwal, Director, NRSC in successfully implementing this work. The support provided by
Sri.K.Abdul Hakeem, Water Resources Division, NRSC and Ms.P.K.Saritha, SDAPSA, NRSC is
also duly acknowledged.
REFERENCES
[1] http://support.esri.com/en/knowledgebase/gisdictionary/term/automatedfeatureextraction
[2] http://pixinsight.com/doc/legacy/LE/19_morphological/median_filter/median_filter.html
[3] Hadi Sadoghi Yazdi, Faranak Homayouni, Impulsive Noise Suppression of Images using Adaptive
Median Filter, Computer Department, Ferdowsi University of Mashhad, Mashhad, Iran
[4] Feature Extraction Technique using Discrete Wavelet Transform for Image Classification. Research
and Development, 2007. Scored 2007. 5th Student Conference, 12-11 Dec. 2007
[5] Ehsan Nadernejad, Edge detection techniques, evaluation and comparisons, Applied Mathematical
Sciences, Vol. 2, 2008, no. 31, 1507 – 1520
[6] Mohamed Ali, David Clausi, Using The Canny Edge Detector for Feature Extraction and
Enhancement of Remote Sensing Images, Systems Design Engineering, University ofWaterloo
[7] S. E. Umbaugh, Computer Vision and Image Processing, Prentice-Hall, Englewood Cliffs, NJ, USA,
1998.
[8] Removal of high density salt and pepper noise in noisy images using decision based unsymmetric
trimmed median filter, International Journal of Computer Trends and Technology - volume 2, Issue 1
- 2011
[9] Zahra Lari, Hamid Ebadi, "Automated Building Extraction from High-Resolution Satellite Imagery
Using Spectral and Structural Information Based on Artificial Neural Networks", ISPRS Hannover
Workshop 2007 High-Resolution Earth Imaging for Geospatial Information, Hannover, Germany,
June 2007.
[10] Mohammadzadeh Tavakoli, A., Valadan Zoej, M.J., Automatic Linear Feature Extraction of Iranian
Roads from High Resolution Multi-spectral Satellite Imagery
[11] Tejaswi Phani K.,Thara Nair, Prasad A.V.V., "Performance Enhancement of DWT Processing using
GPGPU", Proceedings of the Second International Conference on Computer Applications ,2012
[12] NVIDIA.(2010)CUDA ZONE.[Online].http://www.nvidia.com/ object/ cuda_home_new.html
[13] NVIDIA GPU Programming Guide Version 2.5.0 © 2006 by NVDIA Corporation.

15
Authors
Ms. K. Phani Tejaswi did her B.Tech in Electronics and Communications
Engineering,from JNTU College of Engineering, Hyderabad. She is currently working
as Scientist in Data Processing Area in National Remote Sensing Centre, Indian Space
Research Organization (ISRO). She is involved in algorithm development and image
processing in GPGPU.
Mr. D. Shanmukha Rao did his M.Sc, Mathematics from Andhra University,
Visakhapatnam. He is currently working as a Scientist in Data Processing Area in
National Remote Sensing Centre, Indian Space Research Organization (ISRO). He is
involved in algorithm development for data processing and image analysis.
Ms. Thara Nair did her M. Tech, Control Systems Engineering from College of
Engineering, Trivandrum. She worked as Scientist in Vikram Sarabhai Space Centre,
Indian Space Research Organization from 1997 - 2010. She is currently working as
Scientist in Data Processing Area in National Remote Sensing Centre, Indian Space
Research Organization (ISRO). She is involved in High Performance Computing and
algorithm development for image processing applications using GPGPU.
Mr. A.V.V. Prasad did his M.Sc, Physics (Electronics) from Andhra University,
Visakhapatnam in 1985. He is presently the Group Head of Microwave Remote
Sensing and Global Data Processing Group of National Remote Sensing Centre, Indian
Space Research Organization (ISRO). He was involved in the installation of Remote
Sensing satellite data reception systems and test systems like Advanced Front End
Hardware units (AFEH), Serializer systems, Data logging systems etc., for different
satellites like IRS-1C, IRS-P4,IRS-P5, IRS-P6 , TES etc., and a data capturing facility
at Arctic Station, Svalbard and Antarctica.

GPU ACCELERATED AUTOMATED FEATURE EXTRACTION FROM SATELLITE IMAGES

More Related Content

What's hot (20)

Similar to GPU ACCELERATED AUTOMATED FEATURE EXTRACTION FROM SATELLITE IMAGES (20)

Recently uploaded (20)

GPU ACCELERATED AUTOMATED FEATURE EXTRACTION FROM SATELLITE IMAGES