IEEE Geoscience and Remote Sensing Magazine - June 2016 - 86

(a)

(b)

(c)

figure 7. Sky/cloud image segmentation: (a) binary (or two-level) segmentation of a sample input image from the HYTA database,

Average Classification Accuracy

(b) three-level semantic segmentation of a sky/cloud image [73], and (c) probabilistic segmentation of a sky/cloud image [68].

0.9
0.85
0.8
0.75
0.7
0.65
0.6
0.55
0.5
0.45
0.4

PCA

LDA

Using KNN

NCA

No DR

Using SVM

figure 8. The average multiclass classification accuracy using

Heinle features for cloud patch categorization for different methods.
and swarm optimization are the most commonly used unsupervised classification, and (clustering) techniques. Ari and
Aksoy [74] used GMM and particle swarm optimization for
hyperspectral image classification, and Maulik and Saha [75]
used a modified differential evolution-based fuzzy clustering algorithm for satellite images. Such clustering techniques
are also used in ground-based image analysis.
In addition to supervised and unsupervised methods,
semi-supervised learning (SSL) methods are widely used in
remote sensing [76]. SSL uses both labeled and unlabeled
data in its classification framework, helping to create a robust
learning framework that learns the latent marginal distribution of the labels. This is useful in remote sensing, as the availability of labeled data is scarce and manual annotation of data
is expensive. One such example is hyperspectral image classification [77]. In addition to SSL methods, models involving
sparsity and other regularized approaches are also becoming
popular, e.g., Tuia et al. [78] study the use of nonconvex regularization in the context of hyperspectral imaging.
In ground-based image analysis, image classification
refers to categorizing sky/cloud types into various kinds,
e.g., clear sky, patterned clouds, thick dark clouds, thick
88

white clouds, and veil clouds (see the "Dimensionality Reduction" section). To quantify the accuracy of the separation of data in Figure 4, we use several popular clustering
techniques in combination with DR techniques. We use
two classifiers for evaluation purposes, i.e., k-NN and the
support vector machine (SVM). k-NN is a nonparametric
classifier, wherein the output label is estimated using a
majority voting of the labels of a neighborhood. The SVM
is a parametric method that generates a hyperplane, or a
set of hyperplanes, in the vector space by maximizing the
margin between classifiers to the nearest neighbor data.
We evaluate five distinct scenarios, 1) PCA, 2) FA,
3) LDA, 4) NCA, and 5) no DR, and report the classification performances of both k-NN and SVM in each of
these cases. We again use the SWIMCAT [48] database for
evaluation purposes. The training and testing sets consist
of random selections of 50 distinct images, all of which
are downsampled to 32 # 32 pixels for faster computation. Using the 50 training images for each of the categories, we compute the corresponding projection matrix for
PCA, FA, LDA, and NCA. We use the reduced two-dimensional (2-D) Heinle feature for training a k-NN/SVM classifier for scenarios 1-4. We use the original 12-dimensional vector for training the classifier model for scenario
5. In the testing stage, we obtain the projected 2-D feature
points using the computed projection matrix, followed by
a k-NN/SVM classifier for classifying the test images into
individual categories. The average classification accuracies across the five classes are shown in Figure 8.
The k-NN classifier achieves better performance than
the SVM classifier in all of the cases. From the 2-D projected feature space (see Figure 4), it is clear that the data
points belonging to an individual category lie close to
each other. However, it is difficult to separate the different categories using hyperplanes in 2-D space. We observe that the complexity of the linear SVM classifier is
not sufficient to separate the individual classes, as k-NN
performs relatively better in this example. Among the
different DR techniques, LDA and NCA work best with
the k-NN classifier, because these methods also use the
ieee Geoscience and remote sensing magazine

june 2016

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - June 2016