IEEE Geoscience and Remote Sensing Magazine - March 2015 - 24

3.5 Image ClassIfICatIon
Classification is the act of labelling data items to groups
with homogeneous characteristics (classes) based on a given classified training set. The classification taxonomy in
the literature of image processing is quite intricate. The taxonomy is presented from diverse perceptions categorized
on various basis. The main classification is either carried
out on single pixels (pixel-based) or on groups of pixels
(object-based), where groups are assembled through a set
of different segmentation algorithms.
3.5.1 Parametric and non-Parametric
Image classification algorithms can be categorized as either
parametric (ex: maximum likelihood, linear discriminant
analysis) or non-parametric (neural networks, decision
trees, support vector machines and expert systems) [58].
The parametric approach assumes a multivariate normal distribution on the data. The classifier uses the training sample data to define the population statistics (mean
and variance/covariance). Previous literature has discussed
extensively the advantages (robustness and availability)
as well as the main drawbacks of this approach [54]. The
main disadvantages of the parametric approaches are the
assumption of normal distribution (which may not hold
in complex scenes), the ignored mixed-pixels problem,
spectral confusion and the difficult integration of the results with ancillary data [54]. The most popular classifier
in practice is the maximum likelihood.
Non-parametric classifiers, on the contrary, do not assume normal distribution on datasets. Therefore the image
is not classified based on the statistical parameters of the
training sample. Neural network algorithms are the most
widely applied non-parametric classifiers. Previous literature
has suggested its usage in land use/land cover classification
[59]. The main advantages of the non-parametric approach
are: its non-parametric nature itself (i.e. no need to stick to a
pre-defined class model), good performance with noisy inputs, self-adapt capabilities, and fuzzy output values. On the
other hand, the main disadvantages are the problem of possible over-fitting, the long time spent on training of Artificial
Neural Networks (ANNs), and semantic poorness [54, 60].
3.5.2 machine learning
Machine learning (ML) consists of data analysis algorithms
that use the statistical reasoning on the data to develop the
program's own understanding. Machine learning is used
to forecast class memberships based on class characteristics learned from training. Thus, ML has been applied in
image classification to tackle the real-world constraints on
computer vision.
Machine learning can fit into two main categories: supervised and unsupervised learning.
◗ supervised learning: the algorithms are trained on
class-labelled dataset sample
◗ unsupervised learning: the algorithms operate on unlabeled dataset sample
24

In image processing, both the supervised and unsupervised learning algorithms are used to discover groups of
similarity within the data. There are two strategies for the
sampling of training data, which specify whether the classification is supervised or unsupervised.
A) Sampling of training data done by the operator (supervised learning): the chosen set should be from clearly
identified training areas and consider the different
classes for a unbiased representation of the population.
The most popular supervised classifications are maximum likelihood [61], neural networks [62], Support
Vector Machines (SVMs) [63], decision tree [64].
B) Sampling of training data by clustering (unsupervised
learning): in cases of difficulty in identifying clear
training areas by an operator, clustering uses the data
characteristics to create these groups automatically. Using clustering techniques, a set of randomly sampled
data is used to create groups with similar properties
[65]. The clusters are then used for assessing the population statistics. The Iterative Self-Organizing Data Analysis Technique (ISODATA) and K-mean methods are the
most frequently applied methods and they are available
within the most popular image processing software.
3.5.3 rule-based classifiers
Experts interpret the data analysis results with knowledge
from accumulated experience. Machines estimate the classification using a partial expert knowledge. Therefore, with
the different forms of ancillary data, such as road network,
soil map, precipitation, etc., a problem-solving system has
been introduced to develop a knowledge-based classification using the spatial distribution of the patterns. For
example, building density and road network are associated with land-use distribution. The effective use of these
relationships among the produced classes has shown an
improvement in the classification accuracy. This type of
classifiers forms the upcoming trend for its flexibility of
aggregating the contribution of multi-sources data [54,
66]. In that context, e-Cognition software is one of the
most reputable applications of the expert system based on
object-based classification results [67, 68].
3.5.4 ensemble classifiers
Different classifiers perform differently due to the diversity on data types and urban complexity. Each of the classifiers has its strong points and its limitations [69]. Previous
literature has suggested that by appropriately combining
classification outputs from a set of classifiers it is possible
to increase the accuracy of the results compared to the
performance of single classifiers [70, 71]. However, the remaining challenge is to find the appropriate rules of combination for the multi-classification results.
3.6 evaluatIng performanCe
Measuring the classification performance is central in
comparing the classification results from the enormous
ieee Geoscience and remote sensing magazine

march 2015

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - March 2015