See discussions, stats, and author profiles for this publication at: [Link]

net/publication/220827322

Dust Storm Detection Using a Neural Network with Uncertainty and

Ambiguity Output Analysis

Conference Paper · June 2011

DOI: 10.1007/978-3-642-21587-2_33 · Source: DBLP

CITATIONS READS

15 584

4 authors, including:

Mario Chacon Pablo Rivas

Instituto Tecnológico de Chihuahua Baylor University
116 PUBLICATIONS 1,165 CITATIONS 67 PUBLICATIONS 439 CITATIONS

SEE PROFILE SEE PROFILE

Sergio Cabrera
University of Texas at El Paso
128 PUBLICATIONS 788 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Determination of a Vision System for Movement tracking of Clodecera in a Controlled Environment CASCER View project

Ethics Under Pressure View project

All content following this page was uploaded by Mario Chacon on 17 May 2014.

The user has requested enhancement of the downloaded file.

 Dust Storm Detection Using a Neural Network with
Uncertainty and Ambiguity Output Analysis

Mario I. Chacon-Murguía1, Yearim Quezada-Holguín1, Pablo Rivas-Perea2, and

Sergio Cabrera2
1
DSP & Vision Laboratory, Chihuahua Institute of Technology, Chihuahua, Mexico
mchacon@[Link].
2
ECE, University of Texas at El Paso, USA
sergioc@[Link] and privas@[Link]

Abstract. Dust storms are meteorological phenomena that may affect human
life. Therefore, it is of great interest to work towards the development of a
stand-alone dust storm detection system that may help to prevent and/or
counteract its negative effects. This work proposes a dust storm detection
system based on a Artificial Neural Network, ANN. The ANN is designed to
identify not just dust storm areas but also vegetation and soil. The proposed
ANN works on information obtained from multispectral images acquired with
the Moderate Resolution Imaging Spectroradiometer (MODIS) instrument.
Before the multispectral information is fed to the ANN a process to remove
cloud regions from images is performed in order to reduce the computational
burden. A method to mange undefined and ambiguous ANN outputs is also
proposed in the paper which significantly reduces the false positives rate.
Results of this research present a suitable performance at detecting the dust
storm events.

Keywords: Dust storm detection, image segmentation, neural network output

analysis.

1 INTRODUCTION
In recent years weather conditions have attracted the attention of the international
community. Different countries have suffered the consequences of natural phenomena
like earthquakes, tsunamis, floods, hurricane, drought, dust storms, which
significantly affect human life as well as the economy of those countries. These
climate situations have promoted a new interest on remote sensing research because it
offers the potential of better understanding of these phenomena. One of these natural
phenomena that has been studied in recent years is dust storm. Dust storms are
seasonal meteorological phenomena. Usually, dust storms occur in arid and semiarid
regions and can travel long distances. The dust storm may have a negative effect in
our planet and human life in different ways like; cloud formation, respiratory illness,
aerial and terrestrial transportation. This phenomenon can also damage crops and may
cause fertile soil erosion [1]-[3]. Research on dust storm is of great interest because it
can help to find methods to prevent and/or counteract its negative effects. This paper
2 Mario I. Chacon-Murguía1, Yearim Quezada-Holguín1, Pablo Rivas-Perea2, and
Sergio Cabrera2
presents the development of an Artificial Neural Network, ANN, classifier intended to
detect the presence of a dust storm in satellite multispectral images. The paper is
organized in the following sections. Section II presents a literature review of this
topic. The dust storm data used in this work is described in Section III. Analysis of
the classifier is covered in Section IV. Finally the results and conclusions of the
research are commented in Section V.

2 Literature review
In order to establish some aspects of the classifier design, a literature review was
achieved considering the following points: type of instrument used to acquire the
multispectral images, methods used for dust storm detection, performance metrics,
and event location. From 43 papers related to the topic of dust storm and satellite
multispectral images only 13 were considered for the analysis because only these
papers were related specifically to dust storm detection. With respect the type of
sensor used, works reported in [4] to [9] use the Moderate-Resolution Imaging
Spectroradiometer instrument, commonly named MODIS sensor. The researches in
[6] and [7] work with the Aqua MODIS. The sensor AERONET and the AVHRR are
used in [10] and [11] respectively. The work reported in [12] employs the CALIPSO
and in [13] the MERIS sensor is used to detect dust storm in sand regions. The MFR-
7 sensor is mentioned in [14]. Also in [9] it is reported a combination of information
obtained from the sensors MODIS and TOMS. Finally, the MISR sensor is used in
[15]. With regard to the type of image processing techniques used, the information is
the following. All works are based on pixel level feature extraction processing, except
in [9] and [10] where a window feature extraction processing is used instead. The
method used to evaluate the performance of the dust storm detection systems was
only qualitative. The dust storm events reported in the literature were located at;
South Korea, China, Mongolia, India, Egypt, Senegal and East of Africa. In some of
these works it was necessary to adjust some of the parameters of the method
developed in order to obtain correct results.
Considering the information found in the literature review the following points
were determined. It was decided to use the MODIS sensor because it has a good
spectral resolution, 36 bands from 0.62 µm to 14.382 µm. The temporal resolution is
15 minutes and its spatial resolution 1 Km. The image processing technique selected
was pixel level processing because most of the reported works that used this
technique achieved better results than the two papers that used window based
processing.

3 Data Analysis
3.1 Database
The region of interest for dust storm detection is defined as the north region of
Chihuahua State in Mexico and the southwest area of Texas State in the USA. The
dust storm events were acquired from the web page http:// [Link]/
data/ [Link]. These events were acquired by the MODIS sensor. Eight events
were downloaded and their information is given in Table 1.
 Dust Storm Detection Using a Neural Network with Uncertainty and Ambiguity Output
Analysis 3

3.2 Band Selection

Not all the information in the bands of the multispectral images is related to dust
storm events. In order to determine which bands will be used in the design of the
classifier a literature review was achieved. In the works reported in [5]-[9] the bands
B31 and B32 of the MODIS sensor were related to dust storm information. The bands
4 and 5 of the sensor AVHRR were used for dust storm detection in [1] and [2] but
these bands correspond to the bands 31 and 32 of the MODIS sensor. In [6] the band
B29 of the MODIS was incorporated because it provides extra information in clear
days. Based on these evidences we concluded to use the bands B29, B31 and B32.

Table 1 Dust Storm Event Information

Date Hour Date Hour

1. April/6/2001 18:30 hrs. 5. April/15/2003 17:10 hrs.
2. April/10/2001 18:05 hrs. 6. April/15/2003 17:15 hrs.
3. July/2/2002 17:55 hrs. 7. April/15/2003 18:50 hrs.
4. December/17/2002 18:45 hrs. 8. November/22/2003 18:20 hrs.

3.3 Data Selection

In order to obtain reliable samples for the neural network design, data was
statistically selected for each class. The classes considered in this work were dust
storm S, vegetation V, and soil S. The samples selected correspond to samples
obtained from regions of each class that satisfy the following criterion

Dpc  Dc for c  S ,V , G (1)

where Dpc is the distance of the candidate sample pBc of the band B of the region c
to the mean value of the sample mean,  Bnc , of the band B in the region c,

 B1        Bnc  pBnc 

2 2
Dpc  c
 pB1c (2)
Dc is the maximum tolerance distance, one standard deviation  Bnc , of a sample of
a region c to the sample mean of the region.

4 Dust Storm Detection

This section describes the process of the ANN classifier design including a step to
eliminate the clouds present in the images and then it follows with the design of
variations of the ANN classifiers.

4.1 Cloud Elimination

Multispectral images in most of the cases represent huge computational burden,
therefore it is recommended to reduce the amount of information as much as possible.
It was reported in the literature [6], that presence of clouds is an important
perturbation that affects the performance of several dust storm detection methods.
4 Mario I. Chacon-Murguía1, Yearim Quezada-Holguín1, Pablo Rivas-Perea2, and
Sergio Cabrera2

Thus, cloud elimination is a good alternative to reduce the computation load of dust
detection algorithms and at the same time to get rid of possible perturbations. The
work in [4] reports that cloud information can be eliminated using bands B3 and B7.
The work shows that the maximum energy related to clouds is captured by band B3,
meanwhile energy of other elements is minimum in this band. Band B7, contrary to
band B3, capture minimum energy related to clouds and high energy the other
elements. Based on the previous information, it is proposed a difference index ID = B7-
B3 to eliminate cloud information. Using this index a mask image, I M , can be
generated to eliminate the cloud pixels.

0 if I D < 0
IM   (3)
1 if I D > 0

Other important result is that the ID of water pixels will result in zero or near to
zero values. Therefore water pixels will be also eliminated. The application of this
process is illustrated in Figure 1, event #1 April/6/2001. The original information is
shown in Figure 1a, the mask image is shown in Figure 1b and the resulting
information of the event is shown in Figure [Link] cloud elimination process presents
some issues. On one hand, small clouds are not completely eliminated. On the other
hand, some dark vegetation and soil regions are eliminated. However dust storm
information is not significantly affected by this process. An example of this screening
is shown in Figures 1d-f

4.2 Classifiers Design

The classifier was designed to recognize three classes: vegetation, dust storm and
soil; unlike other works that only consider a two-class problem dust storm and not
dust storm. It was decided to include the class vegetation and soil because they may
be of interest in future work.
The structure of the neural network was determined by selecting the best network
in a set of experiments changing the number of layers, 2 and 3 layers, and the number
of neurons in each layer. The best neural network was defined with respect the best
performance associated with the smaller number of layers and neurons by layer. The
ANN model selected for the classifier is a 2-layer Feed-Forward neural network with
15 neurons in the hidden layer, and 3 neuron outputs. The activation functions are
sigmoidal. The training was performed with the backpropagation algorithm using the
Scaled Gradient Algorithm since this algorithm was more effective than other
gradient methods. Among their advantages we can mention that it does not depend on
user parameters like learning rate and momentum. The first classifier, C29, was
designed with the information provided by the bands B29, B31 and B32. The number
of samples used was 5383; 721 samples of vegetation, 1493 of dust storm and 3169 of
soil. The training samples correspond to the 70% of the total number of samples, 15%
for validation and 15% for testing. The performances of this ANN were 95.8%, of
correct classification for training, validation and testing. Table 2 shows the confusion
matrix of the design process. It can be observed that the class dust storm has the best
classification performance considering actual and prediction conditions
 Dust Storm Detection Using a Neural Network with Uncertainty and Ambiguity Output
Analysis 5

a) b) c)

Fig. 1. a) Original image event April/6/2001, b) IM mask, and c) Image without clouds and
water. d), e) and f) zoom-in of the red rectangles in a), b) and c).

Table 2 Confusion matrix of the classifier C29

Vegetation Dust Storm Soil Total

Vegetation 578 0 8 98.6%

Dust Storm 12 1450 20 97.8%
Soil 131 43 3141 94.8%
Total 80.2% 97.1% 99.1% 96%

97.45% followed by Soil, 96.95% and vegetation 89.4%. The total performance is
96%. Figure 2 shows some visual results obtained with this classifier for the events
#2, 5 and 8, corresponding to April/10/2001, April/15/2003 and November/22/2003
respectively. The result of the event April/10/2001 presents some false positives areas
close to the dust storm region. The event April/15/2003 shows a correct detection, but
the event of November/22/2003 involves many false positives regions.

4.3 C29 Output Analysis

The results shown in the previous section led us to investigate the neuron outputs.
The previous results were generated by the following rule Assign pixel pB to class i if
Oi > Oj for i ,j={v,d,s} i ≠ j, where v,d, and s stands for the classes vegetation, dust
storm and soil respectively. This rule assigns a pixel to the class with the higher value
of the neuron outputs. This kind of decision is generally used in some works.
However, this is not a guarantee that a good decision is taken. We can analyze two
hypotheses: the first one supposes that three or two neuron outputs are high but both
with close values; the second hypotheses could be when all outputs are low. In both
cases we can determine a winner output but it does not mean the decision will be
correct. From this analysis we can propose two cases. An undefined case, when the
maximum output of the ANN is less than say 0.6, that is
max(Oi ) < 0.6 for i={v,d,s} (4)
In this situation we can say that the ANN does not have a strong output in any of its
outputs, therefore the class is undefined. The 0.6 threshold was determined by
considering the output of the neurons as an effectiveness class definition percentage,
the value 0.5 represents high vagueness. The second case is when the difference
between two neuron outputs used to make a decision is less than 0.3,
|Oi - Oj| < 0.3 for i, j ={v,d,s} i ≠ j (5)
6 Mario I. Chacon-Murguía1, Yearim Quezada-Holguín1, Pablo Rivas-Perea2, and
Sergio Cabrera2

a) b)
Fig. 2. Results of the classifier C29 for the events: a) April/10/2001, b) November/22/2003.
Cloud /water black, vegetation green, storm yellow, soil brown.
In this circumstance, the outputs of the ANN are so ambiguous that a decision is
not recommended. Using the previous cases two new images can be generated to
analyze the two hypotheses, Iu(x,y) for undefined outputs and IA(x,y) for ambiguous
outputs
Iu(x,y)= {pB(x,y): max(Oi ) < 0.6 for i={v,d,s} } (6)
IA(x,y)= {pB(x,y):|Oi - Oj| < 0.3 for i, j ={v,d,s} i ≠ j } (7)
Figure 3 illustrate the pixels p(x,y) Iu(x,y) and p(x,y) IA (x,y) in white.
Information shown in Figures 3 indicates that many pixels fall in the undefined and
ambiguous cases. Therefore, assigning a pixel to a class under the highest output
criterion is not a recommended criterion. In other analysis it was found that in most
cases pixels corresponding to dust storm regions are well defined.

a) b)
Fig. 3. p(x,y) Iu(x,y) and p(x,y) IA(x,y) of a) April/10/2001, b) November/22/2003.

4.4 C29 ANN Output Adjustment

Based on the analysis described in Section 4.3 a new criterion to determine a
wining neuron was defined. A wining neuron is the neuron with values greater than
0.6 and with a difference greater than 0.3 with respect the other outputs. The new rule
is, Assign pixel pB to class i if Oi >0.6 AND |Oi - Oj| > 0.3 for i ,j={v,d,s} i ≠ j. The
color map to describe the new output of the ANN is as indicated in Table 3. Using
this new criterion the events where classified again. In most of the cases the results
are better because the region of the dust storm is better defined by the new criteria and
false positives are reduced as seen in the blue rectangle in Figure 4. The previous
process confirm that when undefined outputs and ambiguous outputs are analyzed in
the classification processes they positively contribute in the performance of the ANN
classifier without negatively effecting the correct detection of the dust storm region.
 Dust Storm Detection Using a Neural Network with Uncertainty and Ambiguity Output
Analysis 7
Table 3. Color map of the new ANN outputs

Case Description Color

Weak output Less than 0.6
Undefined Outputs with a difference
less than 0.3
Both The two previous cases

a) b)
Fig. 4. First result and result with the new criteria for events a) April/6/2001, b)
April/10/2001.
5 Results and Conclusions
Findings in this research indicate that the cloud elimination method is a good
alternative to reduce the computational burden as well as to get rid of possible
perturbations that may cause a negative effect in the performance of the dust storm
detector. The ANN classifier presents a suitable performance at detecting the dust
storm events. The dust storm events were detected by the method in all events
analyzed in this work except in the event #3 where it was not possible to detect the
event. This event is also complicated to be perceived by a human observer. In regard
to the false detection rate, we can say that it is tolerable because the cost of the event
miss is higher than the false detection error cost. The performance of the ANN
classifier is also acceptable, at least under a visual evaluation, at detecting the other
two classes, vegetation and soil. The proposed method to manage undefined and
ambiguous ANN outputs proved to be an important contribution in this research as the
false positives were significant reduced without causing negatives hits in the dust
storm area.
In conclusion, the proposed dust storm detector system based on ANN can produce
preliminary information related to dust storm detection that may be used for posterior
analysis. Moreover, the statistical selection of a relevant training dataset, allowed the
construction of a low-complexity ANN model. To overcome uncertainty and
ambiguity in ANN outputs, adjustments were made and different criteria were
established, which produced higher accuracy rates, as well as a decrease in false
positives count. However, false positive issues is an important point that needs more
work if the systems is intended to be used as a stand-alone dust storm system.
Acknowledgments. This work was supported by SEP-DGEST, ITCH, UTEP, and
partially supported by CONACYT under grant 193324, SEP-DGRI, and Texas
Instruments Foundation endowed scholarship.
 8 Mario I. Chacon-Murguía1, Yearim Quezada-Holguín1, Pablo Rivas-Perea2, and
Sergio Cabrera2

6 References
1. Janugani Swapna , “Dust Storm Detection Using Classification And Filter Banks”,
M.S. thesis, University Of Texas At El Paso, El Paso, Texas, USA, 2008.
2. Rivera Nancy , “Detection And Characterization Of Dust Source Areas In The
Chihuahuan Desert, Southwestern North America”, M.S. thesis, University Of Texas
At El Paso, El Paso, Texas, USA, 2006.
3. Rivas P., J. Rosiles, and M. I. Chacon M., “Traditional and Neural Probabilistic
Multispectral Image Processing for the Dust Aerosol Detection Problem”, in
Proceedings of the 2010 SSIAI, Austin, TX, May 23-25, pp 169 – 172, 2010.
4. Qu John , Xianjun Hao, Wanting Wang, Lingli Wang and Menas Kafatos, “A Study
of African Dust Storm and Its Effects on Tropical Cyclones over Atlantic Ocean from
Space”, Geoscience and Remote Sensing Symposium, vol. 4, pp 2715-2718, 2005.
5. Hesham M. El-Askary, Sudipta Sarkar, Menas Kafatos, “A Multisensor Approach to
Dust Storm Monitoring Over the Nile Delta”. IEEE transactions on Geoscience and
remote sensing, vol 41, pp 2386-2391, October 2003.
6. Liu San-chao, Liu Qinhuo, Gao Maofang, Chen Liangfu, “Detection of Dust Storms
by Using Daytime and Nighttime Multi-spectral MODIS Images”, IEEE International
Geoscience and Remote Sensing Symposium, pp 294-296, August 2006.
7. Qu John J., Hao Xianjun, Menas Kafatos, and Wan Lingli, “Asian Dust Storm
Monitoring Combining Terra and Aqua MODIS SRB Measurements”. IEEE
Geoscience and Remote Sensing Letters, vol 3, pp 484-486 October 2006.
8. Han Tao, Li Yaohui, Han Hui, Zhang Yongzhong, Wang Yujie, “Automatic
Detection of Dust Storm in the Northwest of China Using Decision Tree Classifier
Based on MODIS Visible Bands Data”, Geoscience and Remote Sensing
Symposium, vol 5, pp 3603-3606, July 2005.
9. Hesham El-Askary, Menas Kafatos, “Potential for Dust Storm Detection Through
Aerosol Radiative Forcing Related to Atmospheric Parameters”, Geosciences and
Remote Sensing Symposium, pp 620 - 623, July 31.
10. Y. J. Kaufman, Didier Tam, 0. Dubovik, A. Karnieli, L. A. Remer, “Absorption of
sunlight by dust as inferred from satellite and ground-based remote sensing”.
Geophysical research letters, vol 28, no 8, pp 1479-1482, April 2001.
11. Xinping Bai, Zhenxin Song, “The Estimation of Dust Aerosol Sources for the
Numerical Simulation of Asian Dust Storms Observed in May 2005 in China”,
Geoscience and Remote Sensing Symposium, 2006
12. Huang, J., P. Minnis, Y. Yi, Q. Tang, X. Wang, Y. Hu, Z. Liu, K. Ayers, C. Trepte,
D. Winker, “Summer dust aerosols detected from CALIPSO over the Tibetan
Plateau”, Geophysical research letters, vol 34, September 2007.
13. Kaiping Wei, Tao Zhang, Bin He, “Detection of Sand and Dust Storms from MERIS
Image Using FE-Otsu Algorithm”, The 2nd International Conference on
Bioinformatics and Biomedical Engineering, 2008. ICBBE 2008.
14. K. O. Ogunjobi, Y. J. Kim, Z. He, “Aerosol optical properties during Asian dust
storm episodes in South Korea”. Theoretical and Applied Climatology, vol. 76, no 1-
2, pp 65-75, Springer Wien, October 2004.
15. El-Askary Hesham, Agarwal Abhishek, El-Ghazawi Tarek, Kafatos Menas, Le-
Moigne Jacquline "Enhancing Dust Storm Detection Using PCA based Data Fusion",
Geoscience and Remote Sensing Symposium, vol. 2, pp 1424 – 1427, July 2005..

View publication stats