=Paper=
{{Paper
|id=Vol-2665/paper17
|storemode=property
|title=Automatic recognition of the number of channels in unidentified multispectral data
|pdfUrl=https://ceur-ws.org/Vol-2665/paper17.pdf
|volume=Vol-2665
|authors=Nina Vinogradova,Andrey Sosnovsky,Natalya Sevastianova
}}
==Automatic recognition of the number of channels in unidentified multispectral data ==
https://ceur-ws.org/Vol-2665/paper17.pdf
Automatic recognition of the number of channels in
unidentified multispectral data
Nina Vinogradova Andrey Sosnovsky Natalya Sevastianova
Department of Radio Electronics and Department of Radio Electronics and Department of Radio Electronics and
Communications Communications Communications
Ural Federal University Ural Federal University Ural Federal University
Yekaterinburg, Russia Yekaterinburg, Russia Yekaterinburg, Russia
n.s.vinogradova@urfu.ru a.v.sosnovsky@urfu.ru n.u.sevastianova@mail.ru
Abstract—The work is devoted to the development of a image presented in the row vector of sequence of bytes of the
method for identifying unknown parameters of multizone source file. Fourier analysis allows us to identify patterns
Earth images got from remote sensing systems. The method alternation of image content, presented in a one-dimensional
allows automatically to determine the method of alternating discrete form, since the Fourier spectrum has sensitivity to
spectral channels and calculate their number for images stored the periodic components of the signal, expressed in the
in files of uncompressed formats. A theoretical justification appearance of peak values of the amplitude component at
based on a change in the shape of the Fourier spectrum with a frequencies corresponding to the relative frequencies of such
change in the method of alternating channels in the data file is components. An analysis of the location of the peaks of the
presented, features of the shape of the spectrum are revealed
Fourier transform can reveal the presence and parameters of
that allow reliable identification of the required characteristics.
The results of applying the algorithm to real Earth images
periodic components in the row vector of the identified file,
from space are presented, its applicability limits are indicated, and then find the multiband image storing formats and the
and recommendations are given for choosing specific number of image channels.
parameters of the algorithm.
II. THE THEORETICAL BASIS OF THE DEVELOPED METHOD
Keywords—remote sensing data, image recognition, At the first step in the development of the algorithm, it is
Fourier analysis, multispectral data necessary to establish general patterns that arise with a
particular multiband image storing. To identify these
I. INTRUDUCTION patterns, four-band test images 100×100 pixels in size were
Nowadays, no modern sphere of human activity can do generated, where each of the bands takes a fixed brightness
without the use of digital images, starting from medicine and value, which is a random number in the range from 0 to 255.
biology and ending with images of the Earth from space. Of The resulting image is laid out in a row vector in three
particular interest are multiband images [1], when each band different ways, corresponding to three different multiband
of a multidimensional image has a certain piece of image storing data: BSQ, BIL and BIP. Typical brightness
information about an object of interest. Such images have profiles of the obtained one-dimensional signals are
gained particular popularity in the field of remote sensing presented in the Fig. 1―3, а. At the next stage, a Fourier
since the use of a combination of different channels allows transform is applied to each of the generated one-
us to obtain a wide range of various derivative characteristics dimensional discrete signals. The results are presented in the
[2,8].
Fig. 1―3, b.
There are three ways of the uncompressed multiband
image storing: BSQ, BIL, and BIP [3]. The BSQ format
stores information into an image file one channel at a time,
wherein, information about each of the channels is
conditionally presented independently of other ones. The
BIL-format supports line by line recording of all channels,
data is stored into the file sequentially row by row. Using the
BIP-format all information is stored into the final file pixel
by pixel, that is, firstly, information is stored in the first pixel
of the first image channel, secondly, the first pixel of the
second image channel and so on. The right choice among
BIL, BIP, and BSQ-formats for multiband data is key to the
success of the correct image opening on an equal basis with
the knowledge of image size (the number of lines and
colons).
Unfortunately, in some cases, for example, when the Fig. 1. The one-dimensional discrete image signal in BSQ format: a)
header file is lost, or if image storage is damaged, it is brightness profile; b) Fourier transform.
impossible to read the image with specialized software
products due to the loss of information about both the image As can be seen from Fig. 1, a brightness profile has a
size and the multiband image storing formats [4]. quasiperiod equal to the number of pixels in the image
Accordingly, it poses the challenge of developing a region. The Fourier image of such an image consists of the
methodology for the automated determination of the two most conspicuous peaks, the first of which is at the
indicated characteristics that would make it possible to origin, and its value is equal to the total brightness of the
subsequently read the data of image correctly. The proposed pixels in the image, the second peak is at the end of the
methodology is based on the Fourier analysis [5] of the
Copyright © 2020 for this paper by its authors.
Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0)
Image Processing and Earth Remote Sensing
coordinates and its value is equal to the amplitude of the III. RECOGNITION ALGORITHM AND ITS ANALYSIS
first harmonic. At the first stage, it is necessary to emphasize the peaks
at the spectrum. In the proposed algorithm, the task is
realized due to block merging, when the spectrum is divided
into N intervals, in each of which the maximum value is
calculated (Fig. 4, a). In the present work, N is set equal to
50, since in the vast majority of remote sensing systems the
number of channels rarely exceeds that value. After that, a
non-recursive averaging filter (n samples) is applied to
smooth the spectrum fluctuations (Fig. 4, b). In the task, n is
set equal to 3, which turns out to be sufficient to smooth out
the existing fluctuations. As n increases, the peak is
excessively blurred, which makes it difficult to detect, at
lower n smoothing does not occur, which can lead to the
appearance of side peaks. At the next stage, the sequence is
converted to binary, for this, it is necessary to choose some
Fig. 2. The one-dimensional discrete image signal in BIL format: a) threshold value, according to which the brightness elements
brightness profile (the first 1000 samples are shown); b) Fourier will be cut off. Since the peaks in the task are strongly
transform. expressed relative to the general background, the median of
the sequence was chosen as the statistics for calculating the
threshold value. The result is shown in the Fig. 4, с. The
flowchart of the first part of algorithm is shown in Fig. 6, a.
Fig. 3. The one-dimensional discrete image signal in BIP format: a)
brightness profile (the first 100 samples are shown); b) Fourier
transform.
The values of the remaining components are
significantly lower than the peak ones and are grouped
around the central one. From Fig. 2 it follows that the vector
line in the BIL format has a periodicity equal to the product Fig. 4. The result of applying the first part of the algorithm to the Fourier
of the number of lines by the number of channels, the quasi image, shown in Figure 3, b: a) The result of maxima searching over
period will be equal to the width of one line. The Fourier N intervals; b) the smoothing result; c) the threshold processing.
transform is similar to the first situation, however, minor
From Fig. 4, c, it follows that the number of channels of
peaks appear on it with an interval equal to the width of the
the image exactly corresponds to the number of intervals
image row, with each k-th peak degenerating to zero, where
with a logical zero. Accordingly, it is necessary to count
k is the number of channels. The most interesting case is
such intervals. Counting is carried out in a cycle during
shown in the Fig. 3. In addition to the first peak with a
which two one-dimensional arrays are formed, each of
height equal to the total brightness of the image, there are
whose elements represents the length of either zero or unit
(k-1) peaks located at frequencies equal to P iM N , i 1, k . interval. The flowchart of the second part of algorithm is
Therefore, if the format for storing multiband data shown in Fig. 6, b. It should be noted that in the spectra of
corresponds to BIP, then by analyzing the Fourier image, real images there may side peaks due to the texture of the
one can find the number of channels of a multi-dimensional terrain, therefore, before the final calculation of the number
image. Thus, it is necessary to develop an algorithm that of channels, such peaks have to be removed. The third part
would detect peaks of the Fourier transform against the of the algorithm, which performs the removal of side peaks,
background of other components of small amplitude. operates as follows: if the array size interval per unit values
corresponds to 1 or 2, then such array intervals per zero
values must be combined. Values equal to 1 and 2 are
selected based on the analysis of real images of the Earth
from space, the principle of operation of the third part of the
VI International Conference on "Information Technology and Nanotechnology" (ITNT-2020) 75
Image Processing and Earth Remote Sensing
algorithm is shown in Fig. 5, The flowchart of the third part TABLE I. SOME STATISTICAL PARAMETERS OF IMAGES
of algorithm is shown in Fig. 6, b. Remote Mean Mean Median
No. of
sensing Image size image spect. spectrum
channels
systems bright. bright. brightness
MODIS/ 2 9200×5416 9685 1.01×1011 2.55×108
250m
ArcGIS 2 1800×3600 465 4.35×108 2.94×107
DEM
SPOT 4 4 3000×3000 144 3.90×108 1.63×107
MODIS/ 5 4600×2708 8567 5.26×1010 3.06×109
500m
Landsat 7 6 512×512 23 1.26×104 1.71×103
AVIRIS 30 350×400 177 4.69×104 6.24×102
Fig. 5. The principle of the third part of the algorithm.
Since the median of the formed sequence acts as a
threshold value during the algorithm operation (Fig. 6a), the
question of the behavior of the median value for different
remote sensing systems is of interest. Table 1 presents the
value of the median and a number of other statistical
parameters of the used images.
As shown in table 1, the median usually turns out to be an
order of magnitude less than the mean value of the spectrum
brightness. Therefore, if the median of the sample is used as
the threshold value, then it can reliably cut off the peaks
from the general background since it always turns out to be
an order of magnitude smaller than the mean value, which is
due to the pronounced quasiperiodicity, especially at BIP and
BIL.
Fig. 6. The flowchart of the algorithm: a) the first part; b) the second part;
с) the third part.
Fig. 7. a) Image fragment obtained by MODIS system (the two red and
IV. THE RESULTS OF EXPERIMENT
one infrared channels are shown); b) Fourier transform of the signal.
The developed algorithm is implemented in MATLAB
18.b [6] and tested using images obtained from various
remote sensing systems, including MODIS (2 and 5
channels) (Fig. 7) [7], SPOT (4 channels) (Fig. 8), Landsat-
7 (6 channels) (Fig. 9) [9], AVIRIS (30 channels) [10], as
well as data of DEM AcrGIS (2 channels). As follows from
the figures, in all cases, the Fourier spectrum of a one-
dimensional signal with the presentation format of
multiband BIP data is divided into the number of intervals
corresponding to the number of channels of the multiband
Fig. 8. a) Image fragment obtained by SPOT-4 system (the red, near
image. For all used images, the algorithm showed the right
infrared and infrared channels are shown);b) Fourier transform of the
operation, except when there was a high percentage of signal.
cloudiness over the image. In this case, there is no
correlation between the different channels of the
multidimensional image, and the quasiperiodic which is
shown in Fig. 1–3 do not arise. It should be noted that such
images usually are not of high value to the researcher, and
the proposed algorithm can be used, including for automated
search of the indicated moments: if for all three types of
storage of multidimensional data, the spectrum of the
Fourier image is divided into one interval, then this indicates
that the brightness values of the channel images are equal to
each other. Fig. 9. a) Image fragment obtained by Landsat-7 system (the red, red edge
and near infrared channels are shown); b) Fourier transform of the
signal.
VI International Conference on "Information Technology and Nanotechnology" (ITNT-2020) 76
Image Processing and Earth Remote Sensing
V. CONCLUSION REFERENCES
A method and algorithm for its implementation has been [1] R.C. Gonzalez and R.E. Woods "Digital Image Processing," London:
Pearson Education, 2017, 1192 p.
developed that automatically calculates the number of
[2] Y.N. Zhuravel and A.A. Fedoseev, "The features of hyperspectral
channels in multiband images stored in files of remote sensing data processing under environment monitoring tasks
uncompressed formats with completely or partially lost solution," Computer Optics, vol. 37, no. 4, pp. 471-476, 2013.
metadata. This algorithm is sufficient for reliable [3] About ArcGIS [Online]. URL: https://www.esri.com/ru-
identification of counting the number of channels into the ru/arcgis/about-arcgis/overview (12.12.2018).
BIP-format, which is one of the most common. The testing [4] L.G. Dorosinskiy and A.A. Kurganski, "Modeling the clutter
of the algorithm was verified on a series of real images of reflection suppression algorithm in synthetic-aperture radar," CEUR
Workshop Proceedings, vol. 1604, pp. 49-57, 2018.
space systems for remote sensing of the Earth, in all cases,
[5] G.B. Folland, "Fourier Analysis and Its Applications," Providence:
the number of channels was determined correctly, and the American Mathematical Society, 2009, 433 p.
applicability limits of the developed algorithm are also [6] MATLAB for Artificial Intelligence [Online]. URL:
indicated. The obtained results are going to form the basis of https://www.mathworks.com (12.12.2018).
the work devoted to the automated recovery of damaged [7] Aqua Earth-observing satellite mission [Online]. URL:
images using any of the three methods of multiband image https://aqua.nasa.gov (12.12.2018).
storing with completely or partially lost metadata based on [8] D.E. Plotnikov, P.A. Kolbudaev and S.A. Bartalev "Identification of
the Fourier spectrum analysis of the byte sequence. dynamically homogeneous areas with time series segmentation of
remote sensing data," Computer Optics, vol. 42, no. 3, pp. 447-456,
2018. DOI: 10.18287/2412-6179-2018-42-3-447-456.
ACKNOWLEDGMENT
[9] USGS Landsat Missions [Online]. URL: https://www.usgs.gov/land-
resources/nli/landsat/landsat-7?qt-science_support_page_related_con
The work was supported by the grant RFFI 19-29- =0#qt-science_support_page_related_con (12.12.2018).
09022\19. [10] AVIRIS: Airborne Visible/Infrared Imaging Spectrometer [Online].
URL: https://aviris.jpl.nasa.gov/ (12.12.2018).
VI International Conference on "Information Technology and Nanotechnology" (ITNT-2020) 77