IEEE Geoscience and Remote Sensing Magazine - December 2021 - 12

and electronic equipment may affect the image quality,
which leads to a certain degree of degradation of the acquired
images. As presented in Figure 3, these images
cannot fully meet the requirements of precise interpretation
in real-world applications. There are two main
categories of factors that degrade the image quality. The
first one is the factors that possibly appear in the imaging
process, such as noise, blurring, cloud occlusion, missing
information, shadow, and so on. These kinds of factors
are the main reason for remote sensing image degradation.
Another category of factors arises from the limitations
of sensor production technologies and application
scenarios. Because spectral, spatial, and temporal resolutions
are often mutually restricted, imaging sensors can
achieve high resolution in only one of these three aspects.
For these kinds of low-quality images, some methods for
improving the image quality should be applied. Multisource
satellite data with different resolutions should be
complementary to obtain the required data.
2) Object variations: An HRRS image can cover an extensive
area of the Earth's surface and contain many kinds
of objects. The scale variations of object instances in
HRRS images are great, and some objects are very
small. As depicted in Figure 4, some objects always take
a small proportion of a total image and show weak feature
response; for example, the width of a small ship
can be fewer than 25 pixels. Different resolution, scale,
color, shape, and texture changes residing within a
single category create high intraclass variations for objects.
These kinds of small-object instances may likely
crowd into a specific region of aerial images. Additionally,
HRRS images are noisy, and the features of objects
easily change when affected by weather, illumination,
and occlusions. Some specific targets are adversarial
and camouflaged, making them difficult to identify
effectively. Another critical problem is that the annotation
samples may be insufficient. At present, there are
more than 2,000 satellites in orbit around the world;
they generate more than a petabyte of data every day.
However, there are roughly only 100 GB of annotated
data for target detection.
3) Complex context: Generally, the background and context
of objects of interest are complex and crowded with other
type of objects, as displayed in Figure 5. Natural images
are often taken from horizontal perspectives, while HRRS
images are typically taken as bird's-eye views; this implies
that many objects of interest form complex spatial patterns
with the background. The intricate patterns increase
the difficulty of object detection in HRRS images.
Considering the three aspects of the challenges, remote
sensing small, weak objects can be defined as 1) data quality,
i.e., HRRS images for small, weak object detection may be
of low quality due to the noise, illumination change, occlusion,
and so forth introduced in the imaging process; (2) objects,
i.e., they are of small scale, have weak feature response
with many categories showing high intraclass variations and
a nonuniform distribution, and may lack annotation samples;
(3) context, i.e., the context is complex and changeable,
(a)
(b)
(c)
IKONOS
(0.8-1 m)
WorldView-3
(0.31 m)
(d)
(e)
(f)
FIGURE 3. Some problems caused by image quality: (a) blur, (b) noise, (c) missing information (d) cloud, (e) shadow, and (f) multisource data.
12
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE DECEMBER 2021

IEEE Geoscience and Remote Sensing Magazine - December 2021

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - December 2021