DEVELOPMENT AND INNOVATIONS IN SCIENCE
International scientific-online conference
134
THE POWER OF REMOTE SENSING CLASSIFICATION FOR
ENVIRONMENTAL MONITORING AND MANAGEMENT
Kattayev Bobir Sobirovich
“TIIAME” National Research University, 2nd year graduate student in
Environmental Sciences (GIS).
https://doi.org/10.5281/zenodo.11353308
Abstract:
Remote sensing, the science of acquiring information about Earth's surface
from a distance, has become a critical tool for environmental monitoring and
management. Extracting meaningful insights from this vast amount of data
hinges on the power of remote sensing classification. This technique categorizes
pixels within an image based on their spectral characteristics, allowing scientists
to identify and map various features on the Earth's surface. This article explores
the capabilities of remote sensing classification in environmental applications.
We will examine how classification facilitates the monitoring of land cover
change, deforestation, forest health, water quality, and other critical
environmental parameters. By enabling the creation of detailed and up-to-date
maps, remote sensing classification empowers environmental managers to track
changes, assess risks, and develop effective strategies for sustainable resource
management.
Key words:
Remote Sensing, Classification, monitoring, pixel.
Introduction
The overall objective of image classification procedures is to automatically
cate-gorize all pixels in an image into land cover classes or themes (
Lillesand et.
al., 2015, p 537
).
Supervised classification is the procedure most often used for quantitative
analysis of remote sensing image data (
Richards et. al., 2006, p 193
).
In classification based on prototypes, the objective is to make the features
so unique and easily detectable that classification itself becomes a simple task
(
Gonzalez et.al.,2018, p 904
).
Image classification is an important part of the fields of remote sensing,
image analysis, and pattern recognition (
Campbell et.al., 2011, p 335
).
Digital image classification uses the spectral information represented by
the digital numbers in one or more spectral bands, and attempts to classify each
individual pixel based on this spectral information (
Canada Centre for Remote
Sensing 2020, p161
).
DEVELOPMENT AND INNOVATIONS IN SCIENCE
International scientific-online conference
135
Classification of images helps to study the information on the ground more
accurately.
Supervised Classification
We use a hypothetical example to facilitate our discussion of supervised
classifi cation. In this example, let us assume that we are dealing with the
analysis of five-channel airborne multispectral sensor data. (The identical
procedures would apply to Landsat, SPOT, WorldView-2, or virtually any other
source of multispectral data.) Figure 1. shows the location of a single line of data
col-lected for our hypothetical example over a landscape composed of several
cover types. For each of the pixels shown along this line, the sensor has
measured scene radiance in terms of DNs recorded in each of the five spectral
bands of sen-sing: blue, green, red, near infrared, and thermal infrared. Below
the scan line, typical DNs measured over six different land cover types are
shown. The vertical bars indicate the relative gray values in each spectral band.
These five outputs represent a coarse description of the spectral response
patterns of the various terrain features along the scan line. If these spectral
patterns are sufficiently dis-tinct for each feature type, they may form the basis
for image classification (
Lillesand
et. al., 2015,
p 539
).
Figure 1.
Figure 7.34 Selected multispectral sensor measurements made along one
scan line. Sensor covers the following spectral bands: 1, blue; 2, green; 3, red; 4,
near infrared; 5, thermal infrared.
The Classification Stage
Numerous mathematical approaches to spectral pattern recognition have
been developed. Our discussion only scratches the surface of this topic.We
illustrate the various classification approaches with a two-channel (bands 3 and
4) subset of our hypothetical five-channel multispectral sensor data set. Rarely
DEVELOPMENT AND INNOVATIONS IN SCIENCE
International scientific-online conference
136
are just two channels employed in an analysis, yet this limitation simplifies the
graphic portrayal of the various techniques. When implemented numerically,
these procedures may be applied to any number of channels of data. Let us
assume that we take a sample of pixel observations from our two-channel digital
image data set. The two-dimensional digital values, or measurement vectors,
attributed to each pixel may be expressed graphically by plotting them on a
scatter diagram (or scatter plot), as shown in Figure 2. In this diagram, the band
3 DNs have been plotted on the y axis and the band 4 DNs on the x axis. These
two DNs locate each pixel value in the two-dimensional “mea-surement space”
of the graph. Thus, if the band 4 DN for a pixel is 10 and the band 3 DN for the
same pixel is 68, the measurement vector for this pixel is represented by a point
plotted at coordinate (10, 68) in the measurement space.
Let us also assume that
the pixel observations shown in Figure 2. are from areas of known cover type
(that is, from selected training sites). Each pixel value has been plotted on the
scatter diagram with a letter indicating the category to which it is known to
belong. Note that the pixels within each class do not have a single, repeated
spectral value. Rather, they illustrate the natural centralizing tendency yet
variability of the spectral properties found within each cover class. These
“clouds of points” represent multidimensional descriptions of the spectral
response patterns of each category of cover type to be interpreted. The following
classification strategies use these “training set” descriptions of the category
spectral response patterns as interpretation keys by which pixels of unidentified
cover type are categorized into their appropriate classes (
Lillesand
et. al., 2015,
p 541
).
DEVELOPMENT AND INNOVATIONS IN SCIENCE
International scientific-online conference
137
Figure 2. Pixel observations from selected training sites plotted on scatter
diagram.
Minimum-Distance-to-Means Classifier
Figure 3. illustrates one of the simpler classification strategies that may be
used. First, the mean, or average, spectral value in each band for each category is
determined. These values comprise the mean vector for each category. The
category means are indicated by þsymbols in Figure 3. By considering the two-
channel pixel values as positional coordinates (as they are portrayed in the
scatter diagram), a pixel of unknown identity may be classified by computing the
distance between the value of the unknown pixel and each of the category
means. In Figure 3, an unknown pixel value has been plotted at point 1. The
distance between this pixel value and each category mean value is illustrated by
the dashed lines. After computing the distances, the unknown pixel is assigned
to the “closest” class, in this case “corn.” If the pixel is farther than an analyst-
defined distance from any category mean, it would be classified as “unknown.”
The minimum-distance-to-means strategy is mathematically simple and
computationally efficient, but it has certain limitations. Most importantly, it is
insensitive to different degrees of variance in the spectral response data. In
Figure 3, the pixel value plotted at point 2 would be assigned by the distance-to-
means classifier to the “sand” category, in spite of the fact that the greater varia-
bility in the “urban” category suggests that “urban” would be a more appropriate
class assignment. Because of such problems, this classifier is not widely used in
applications where spectral classes are close to one another in the measurement
space and have high variance
(
Lillesand
et. al., 2015,
p 542
).
DEVELOPMENT AND INNOVATIONS IN SCIENCE
International scientific-online conference
138
Figure 3.
Minimum distance to means classification strategy.
Parallelepiped Classifier
We can introduce sensitivity to category variance by considering the range
of values in each category training set. This range may be defined by the highest
and lowest digital number values in each band and appears as a rectangular area
in our two-channel scatter diagram, as shown in Figure 4. An unknown pixel is
classified according to the category range, or decision region, in which it lies or
as “unknown” if it lies outside all regions. The multidimensional analogs of these
rectangular areas are called parallelepipeds, and this classification strategy is
refer-red to by that tongue-twisting name. The parallelepiped classifier is also
very fast and efficient computationally. The sensitivity of the parallelepiped
classifier to category variance is exempli-fied by the smaller decision region
defined for the highly repeatable “sand” cate-gory than for the more variable
“urban” class. Because of this, pixel 2 would be appropriately classified as
“urban.” However, difficulties are encountered when category ranges overlap.
Unknown pixel observations that occur in the overlap areas will be classified as
“not sure” or be arbitrarily placed in one of the two overlapping classes. Overlap
is caused largely because category distributions exhibiting correlation or high
covariance are poorly described by the rectangular decision regions. Covariance
is the tendency of spectral values to vary similarly in two bands, resulting in
elongated, slanted clouds of observations on the scatter diagram. In our
example, the “corn” and “hay” categories have positive covar-iance (they slant
upward to the right), meaning that high values in band 3 are generally
associated with high values in band 4, and low values in band 3 are associated
with low values in band 4. The water category in our example exhibits negative
covariance (its distribution slants down to the right), meaning that high values
DEVELOPMENT AND INNOVATIONS IN SCIENCE
International scientific-online conference
139
in band 3 are associated with low values in band 4. The “urban” class shows a
lack of covariance, resulting in a nearly circular distribution on the scat-ter
diagram
(
Lillesand
et. al., 2015,
p 543
).
Figure 4. Parallelepiped classification strategy
Gaussian Maximum Likelihood Classifier
The maximum likelihood classifier quantitatively evaluates both the
variance and covariance of the category spectral response patterns when
classifying an unknown pixel. To do this, an assumption is made that the
distribution of the cloud of points forming the category training data is Gaussian
(normally dis-tributed). This assumption of normality is generally reasonable
for common spec-tral response distributions. Under this assumption, the
distribution of a category response pattern can be completely described by the
mean vector and the covar-iance matrix. Given these parameters, we may
compute the statistical probability of a given pixel value being a member of a
particular land cover class. Figure 5, shows the probability values plotted in a
three-dimensional graph. The vertical axis is associated with the probability of a
pixel value being a member of one of the classes. The resulting bell-shaped
surfaces are called probability density func-tions, and there is one such function
for each spectral category (
Lillesand
et. al., 2015,
p 544
).
DEVELOPMENT AND INNOVATIONS IN SCIENCE
International scientific-online conference
140
Figure 5.
Probability density functions defined by a maximum likelihood
classifier.
The probability density functions are used to classify an unidentified pixel
by computing the probability of the pixel value belonging to each category. That
is, the computer would calculate the probability of the pixel value occurring in
the distribution of class “corn,” then the likelihood of its occurring in class
“sand,” and so on. After evaluating the probability in each category, the pixel
would be assigned to the most likely class (highest probability value) or be
labeled “unknown” if the probability values are all below a threshold set by the
analyst.
In essence, the maximum likelihood classifier delineates ellipsoidal
“equi-probability contours” in the scatter diagram. These decision regions are
shown in Figure 6. The shape of the equiprobability contours expresses the
sensitivity of the likelihood classifier to covariance. For example, because of this
sensitivity, it can be seen that pixel 1 would be appropriately assigned to the
“corn” category.
An extension of the maximum likelihood approach is the
Bayesian classifier. This technique applies two weighting factors to the
probability estimate. First, the analyst determines the “a priori probability,” or
the anticipated likelihood of occurrence for each class in the given scene. For
example, when classifying a pixel, the probability of the rarely occurring “sand”
category might be weighted lightly, and the more likely “urban” class weighted
heavily. Second, a weight asso-ciated with the “cost” of misclassification is
applied to each class. Together, these factors act to minimize the “cost” of
misclassifications, resulting in a theoretically optimum classification. In practice,
most maximum likelihood classification is performed assuming equal
probability of occurrence and cost of misclassification for all classes. If suitable
DEVELOPMENT AND INNOVATIONS IN SCIENCE
International scientific-online conference
141
data exist for these factors, the Bayesian implementation of the classifier is
preferable (
Lillesand
et. al., 2015,
p 546
).
Figure 6. Equiprobability contours defined by a maximum likelihood classifier.
Conclusion
Image classification is a powerful technique with myriad applications across
various domains, from healthcare to security and beyond. Through our
exploration, we’ve witnessed its ability to accurately categorize and interpret
visual data, enabling automation and efficiency in tasks that were once labor-
intensive. While significant progress has been made, challenges such as
robustness to noise, scalability, and interpretability persist. Continued research
and development in areas like deep learning architectures, data augmentation,
and transfer learning promise to address these challenges and push the
boundaries of what’s achievable. As we move forward, it’s crucial to not only
refine the accuracy and speed of classification algorithms but also to ensure
ethical considerations are integrated into their deployment, fostering trust and
accountability. Ultimately, image classification stands as a testament to the
transformative potential of machine learning in reshaping how we interact with
and derive insights from visual data.
References:
1. Thomas M. Lillesand, Ralph W. Kiefer, Ralph W. Kiefer 2015 “Remote sensing
and image interpretation” Seventh Edition ISBN 978-1-118-34328-9
DEVELOPMENT AND INNOVATIONS IN SCIENCE
International scientific-online conference
142
2. John A. Richards, Xiuping Jia 2006 “Remote Sensing Digital Image Analysis”
Fourth Edition ISBN-10 3-540-25128-6
3. Rafael C. Gonzalez, Richard E. Woods 2018 “Digital Image Processing” Fourth
Edition ISBN 10: 1-292-22304-9
4. James B. Campbell, Dorothy Ann Campbell 2011“Introduction to Remote
Sensing” Fifth edition ISBN 978-1-60918-176-5
5. Canada Centre for Remote Sensing 2020 “Fundamentals of Remote Sensing”
