IEEE Geoscience and Remote Sensing Magazine - September 2019 - 82

TABLE 4. THE ADVANTAGES AND DISADVANTAGES OF VARIOUS METHODS IN ACCORDANCE WITH RSIs' SPATIAL RESOLUTION.
SPATIAL
DATA
RESOLUTION SOURCE
Coarse
resolution
(>100 m)

Medium
resolution
(10-100 m)

High
resolution
(<10 m)

AVHRR
MODIS
DMSP/
OLS

Landsat
(MSS, TM,
ETM+,
OLI/TIRS)
ASTER
TerraSAR-X
ENVISAT
Nighttime
light data

IKONOS
WorldView
QuickBird
ALOS/
SPOT
Lidar data
(1:250-
1:10,000)

CATEGORY
Subpixel
based

Subpixel
based;
pixel
based

pixelbased/
objectedbased

MAJOR
METHODS

ADVANTAGES

DISADVANTAGES

REFERENCE

Regression

Simple operation and clear correlation

Unpredictability of variables;
dependence on historical data

[143]
[208]

Threshold

Simple operation

Difficult definiteness threshold

[202]

SMA

Effective handling of mixed-pixels
problem with low cost

Over- or underestimated and
lower water surface identification

[101], [209]
[210], [12]

Linear SMA

Based on subpixel classification with
higher efficiency

Suitability for simple landscapes;
over- or underestimation in some
regional landscapes

[109], [211]
[24]

SMA/linear
SMA

Based on subpixel classification with
higher efficiency, low cost

Over- or underestimation in some
regional landscapes

[212], [108]
[77], [128]

Principal
component
analysis

Unsupervised learning without subjecting to sample labels; effective dimension
reduction function; simple calculation

A linear algorithm for complex
polynomials (e.g., hyperspectral
images); weak performance

[158]
[213]

ANNs

Good fault tolerance and robustness;
parallel distribution processing and
self-learning function

Dependence on training sample
quality and cover degree; higher
space and time complexity

[123]

SVM

Conversion of multidimensional data
into higher linearly separable dimensional space data and construction of
optimal hyperplane

Storage of a certain amount
of training samples and kernel
matrixes; higher space and time
complexity

[10], [139]
[138]

RF

Higher classification accuracy; faster
training; ability to process high-dimension and multifeature data

Overfitting of classification and regression of some data with larger
noise; high human disturbance

[214], [53]
[215]

Regression

Quick and easy modeling; suitability
for not very complex and not largescale data

Dependence of accuracy on a
priori knowledge

[81], [5],
[126]
[208], [209]

CART

Effective reduction of the complexity of
the nonlinear relationships among the
input variables; low time complexity

Limitation of spatial resolution
and heterogeneity of landscapes;
inability to achieve cartographic
accuracy requirements

[143]
[144]
[50]

NDISI/IBI/BCI/
MNDISI

Low-cost computing; higher efficiency
and sensitive spectral response

Limitation of high contrast among
the selected spectral bands

[172], [173]
[177], [179],
[176], [216]

Expert systems

Provision of potential knowledge-based
classification with more mature rules;
higher execution efficiency

Dependence of classification accuracy on a priori knowledge and certain
rules with absenting self-learning

[57]
[136]

Multialgorithm
fusion

Each algorithm complementarity; acting
out better result

Higher space and time complexity;
indefinite algorithm selection

[158], [124]
[153], [217]

SMA/linear
SMA

Reduction or affecting of mixed pixels
and boundary

Lower execution efficiency in
high-resolution RSIs

[126]

SVM

Controllable classification result with
known training samplings

Stronger anthropogenic interference
in selection of training samplings

[139]
[218]

k-means

Higher action; efficient and lower
space and time cost

Higher space and time complexity;
lower clustering classification
accuracy with k hard to estimate

[149]
[128]

ANNs

Organizing and product training
samples; stronger self-learning; acting
out well

Difficulty of acquiring the data
overall, leading to missing data

[108], [76]
[150]

RF

Parallel processing with faster training
speed

Overdrawing in some noise pollution images or regression problems

[140], [141]
[165]

Genetic
algorithm

Better accuracy, with strong robustness
and method expandability

High space and time complexity

[142]

NDII/BCI

Higher execution efficiency based on
band math

Limitation of high contrast among
the selected spectral bands

[177]
[174], [219]

Multialgorithm
fusion

Improving classification accuracy
with advantages of complementary
multialgorithms

Higher space and time complexity
and algorithm selection

[37], [161]
[213], [217]

Deep learning
(CNNs/DBN/
CDBNs)

Multifeature extraction and unsupervised learning; higher classification
accuracy

Higher space and time complexity;
local-scale UIS mapping

[166], [168]
[167], [165]
[171]
(continued)

82

IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE

SEPTEMBER 2019
IEEE Geoscience and Remote Sensing Magazine - September 2019

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