IEEE Geoscience and Remote Sensing Magazine - September 2019 - 73

ENDMEMBER SELECTION PROBLEM
Endmember selection seriously affects the decomposition
accuracy of mixed pixels. In actual urban scenes, because of
the complexity of ground object types, endmember selection is more subjective. The spectral heterogeneity of UIS is
also difficult to distinguish, given the limitations of medium to low spatial-resolution images. Endmember selection
is especially difficult in global UIS detection [112].
OVER- OR UNDERESTIMATION PROBLEM
The influence of ground objects' complexity on mixed pixels
differs. SMA-based methods may lead to underestimation
of UIS in areas with sparse impervious surfaces (e.g., rural-
urban fringe areas) and overestimation in areas with dense
buildings (e.g., commercial districts and city centers) [113].
LOW-REFLECTION GROUND OBJECT DISTURBANCE
The spectral reflection values of water and shadows are low
and similar to LIS spectral reflection values, as mentioned
previously; thus, these objects are easily confused with LIS.
To reduce confusion with low-reflection ground objects
(e.g., water and shadows), SMA-based methods often remove these disturbances using mask preprocessing [114].
Yet such preprocessing affects the accuracy of UIS detection.

SMA Methods

MISSING SPATIAL DISTRIBUTION INFORMATION
SMA-based techniques can extract the abundance value
of pixels of all types of ground objects, but they cannot
infer the specific spatial distribution of ground objects in
the pixels. SMA-based methods lead to a lack of detailed
spatial information in medium to low spatial-resolution
RSIs [115]. This approach can be combined with other
methods to improve UIS detection accuracy (e.g., Urban
Index techniques and spatial feature extraction methods).
IMAGE CLASSIFICATION METHODS
The extraction of various features from RSIs is related to
resource data and algorithm selection [116]. In the following discussion, image classification approaches are divided
into four categories according to the scale of pixel processing: techniques that are subpixel based, pixel based, object
based, and deep-learning based. These algorithms tend to
be general, although some classification algorithms can be
applied in other types of classification.
SUBPIXEL-BASED METHODS
Procedures that are subpixel based divide mixed pixels into
independent attribute units according to proportions with
the spatial information in an original low-resolution image

Linear SMA Model

SMA Based

Nonlinear SMA Model

Subpixel-Based Classifier

Categories of Methods
and Models of UIS Detection

Linear Mixture Modeling
Sub-Pixel-Based Methods
Pixel Segmentation
Image
Classification Methods

Pixel-Based Methods

Object-Based Methods
Deep-Learning Methods
Urban
Index Methods

Supervised
Classification Methods

SVM/DTs/RF/
ANNs/Regression
Analysis Methods ...

Unsupervised
Classification Methods

k-Means/SelfOrganizing Maps...

CNNs/DBN/DBMs/CDBNs ...

NDISIv/UI/NDII/IBI/NDBI/B
CI/MNDISI...
Spatial-Based Fusion

Multisource Data Fusion
Models

Spatial-Spectral-Based
Fusion
Spatiotemporal-Based
Fusion

FIGURE 8. The generalized classification of major UIS detection algorithms. NDISI: Normalized Difference Impervious Surface Index; UI:

Urban Index; NDII: Normalized Difference Impervious Index; IBI: Index-Based Built-up Index; NDBI: Normalized Difference Built-up Index;
BCI: Biophysical Composition Index; MNDISI: Modified Normalized Difference Impervious Surface Index; SVM: support vector machine;
RF: random forest; CNNs: convolutional neural networks; DBMs: deep Boltzmann machines.
SEPTEMBER 2019

IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE

IEEE Geoscience and Remote Sensing Magazine - September 2019

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