IEEE Geoscience and Remote Sensing Magazine - September 2019 - 75

operations, and ground cover classification with high UIS
extraction accuracy [152]. Moreover, intelligent learning
algorithms will be used in UIS automatic detection and
mapping based on deep-learning mechanisms in the near
future (see the "Evaluation" section).
As an example, we verified a typical supervised algorithm and unsupervised algorithm. Figure 9 illustrates
land cover classification over part of the Purple Mountains
in Nanjing, China, using 4-m-resolution IKONOS multispectral images in tandem with SVM, ANNs, DTs, RF, and
k-means algorithms. In particular, the red represents UIS
in Figure 9. The k-means approach is an unsupervised classification method, whereas the other four are supervised
methods using five regions of interest (i.e., impervious
surfaces, forest, grassland, bare soil, and water). The classification results reveal that the supervised methods (SVM,
ANNs, DTs, and RF) demonstrated a higher producer accuracy of approximately 97% ± 0.11% for impervious surfaces and a higher producer accuracy of around 94% ± 1.7%
for grassland. However, the producer accuracy rose to 99%
for forest because more samples were selected in that area.
The unsupervised k-means method exhibited a higher
commission and omission rate due to misclassification,
with over- or underestimation of impervious surfaces and

method involves disparities in UIS coverage of images during different seasonal phases. In seasons when vegetation is
flourishing, plant cover may be overestimated, and the corresponding UIS may be underestimated. On the other hand,
when deciduous vegetation has shed its leaves, plant cover
may be underestimated and UIS overestimated [147].
UNSUPERVISED CLASSIFICATION METHODS
Unsupervised classification is seldom affected by human
intervention during the classification process, has no need
for known samples, and possesses a lower operating cost
[148]. Representative unsupervised classification algorithms
include k-means [149], self-organizing maps [150], and hierarchical clustering methods [151]. In a distance-based kmeans algorithm, distance is used as an evaluation index of
similarity. However, the clustering number (e.g., k) is difficult to estimate beforehand, and over- and underestimation
of clustering numbers will reduce clustering accuracy. Unsupervised classification algorithms are, therefore, rarely used
in UIS detection because of their low classification accuracy.
Literature on unsupervised UIS detection via the development of deep-learning algorithms [81] is scarce. Deep
learning can reveal multiple RSI features and perform
self-training and self-verification, multiple convolution

(a)

(b)

(c)

(d)

(e)

(f)

Impervious Surface

Forest

Grassland

Bare Soil

Water

FIGURE 9. The land cover classifications in the Purple Mountains region of Nanjing, China, derived from a 2013 IKONOS image. (a) An
IKONOS natural color composite (red: band 3; green: band 2; and blue: band 1). The classifications results using the (b) SVM, (c) ANN, (d)
DT, (e) RF, and (f) k-means methods.
SEPTEMBER 2019

IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE

75
IEEE Geoscience and Remote Sensing Magazine - September 2019

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - September 2019
