Traditional and Neural Probabilistic Multispectral Image Processing for the Dust
Aerosol Detection Problem
P. Rivas-Perea, J. G. Rosiles
Department of Electrical and Computer Engineering
The University of Texas El Paso
El Paso, TX, USA
privas@miner.utep.edu, grosiles@utep.edu
M. I. Chacon M.
Graduate Studies Department
Chihuahua Institute of Technology
Chihuahua, Chih., Mexico
mchacon@ieee.org
Abstract—This paper address the dust aerosol detection
problem based on a probabilistic multispectral image analysis.
Two classifiers are designed. First the Maximum Likelihood
classifier is adapted to mode different types of atmospheric
components. The second is a Probabilistic Neural Network
(PNN) model. The data sets are MODIS multispectral bands
from NASA Terra satellite. Findings indicate that the PNN
presents a better classification performance than the ML
classifier using manual segmentations as ground truth. The
proposed algorithm is capable of real-time processing at 1 km
resolutions which is an improvement compared to the 10 km
resolution currently provided by other approaches.
Keywords-Maximum likelihood classification; Neural net-
works; Image processing; Remote sensing.
I. INTRODUCTION
Advances in remote sensing like multispectral instruments
allow imaging of atmospheric and earth materials based on
their spectral signature over the optical range. In particular,
dust air-borne particles (aerosols) propagated through the
atmosphere in the form of dust storms can be detected
through current remote sensing instruments. Dust aerosols
are a major cause of health, environmental, and economical
hazards, and can adversely impact urban areas [1]. From
a scientific perspective, understanding dust storm genesis,
formation, propagation and composition is important to
reduce their impact or predict their effect (e.g., increase of
asthma cases).
Several methods for dust aerosol detection exist [2]. Some
of the most relevant systems are based in the Moder-
ate Resolution Spectroradiometer (MODIS) Aerosol Optical
Thickness (AOT) product [3] which is provided by the
NASA Terra satellite. However, AOT products require a
considerable amount of processing that introduces a sig-
nificant delay (i.e., two days after satellite pass) before it
can provide useful information on aerosol events. Other
approaches are based on the so-called ”band-math” [1]
where simple operations between bands are used to provide
a visual (and subjective) display of the presence of dust
storms.
Given the large amounts of data produced by the MODIS
instrument, it is also desirable to have automated systems
that assists scientist on finding or classifying different earth
phenomena. For example, Aksoy, et al. [4], developed a
visual grammar scheme that integrates low-level features
to provide a high level spatial scene description on land
cover and land usage. As far as the authors know, similar
automated schemes for dust detection based on statistical
pattern recognition techniques have not been reported.
In this paper we present two methods for the detection
of dust storms from multispectral imagery using statistical
classifiers. Based on reported data, we present a feature set
that allows high performance, accuracy, and real-time detec-
tion of dust aerosol. The proposed feature set is extracted
from MODIS spectral bands and tested with the maximum
likelihood classifier and the probabilistic neural network
(PNN). We will show that the PNN approach provides a
better detection and representation of dust storm events.
This paper is organized as follows. Section 2 of the paper
introduces the dust aerosol multispectral analysis. The ML
and PNN models are explained in Section 3 and 4. Section 5
presents experimental results, followed by a brief discussion
on the proposed schemes. Finally, conclusions are drawn in
Section 6.
II. S
ELECTION AND ANA LY S I S O F SPECTRAL BANDS
The MODIS instrument is part of NASA Terra satellite.
MODIS data is currently used in the analysis of different
phenomena like sea temperature and surface reflectivity.
MODIS provides information in 36 spectral bands between
wavelengths 405nm and 14.385μm. These bands are avail-
able in MODIS Level 1B file organization. In the case of dust
aerosol, visual assessment can be achieved using MODIS
bands B1, B3, and B4 which correspond to the range of
human visual perception [5]. An RGB composite true color
image can be produced by the mapping R = B1, G = B4,
and B = B3. Hao et al. [6] demonstrated that bands
B20,B29,B31 and B32 can also be utilized for dust aerosol
visualization. Ackerman et al. [7] demonstrated that band
subtraction B32 − B31 improves dust storm visualization
contrast. Based on these findings, we will form feature