IEEE Geoscience and Remote Sensing Magazine - December 2021 - 9

In the past 20 years, the increasing image interpretation
accuracy of these applications has enabled them to meet
the requirements needed in actual scenarios and thus significantly
promotes the development of Earth observation
technologies and object-detection approaches. The spatial,
temporal, and spectral resolutions of Earth observation
sensors have also been greatly improved [1]-[3]. For
instance, the images from Google Earth (Google Inc.) [4]
have resolutions of up to approximately 0.5 m. WorldView-3
(DigitalGlobe, Inc.) [5] provides a 0.31-m panchromatic resolution
and a 1.24-m multispectral resolution. These HRRS
images show more texture and shape and additional detailed
information about geospatial objects as well as complex
spatial patterns. The data volume of HRRS images has
also dramatically increased, and a massive number of images
is now accessible. The advantages of HRRS images are
that they can offer the most economical and efficient way
to achieve full-time, high-precision Earth surface monitoring
with global coverage and fast detection of small-scale
(small), weak-feature-response (weak), and nonuniformly
(sparsely or densely) distributed objects is of great significance
when meeting the requirements of real scenarios in
many special applications, such as military reconnaissance,
national security, urban monitoring, and geological disaster
monitoring.
Unlike natural images, which are often clearer and contain
several categories of objects, HRRS images cover an
extensive range of the Earth's surface and involve a massive
number of objects. The objects vary in scale, color, shape,
and texture; their features are easily affected by weather,
illumination, occlusion, and imaging conditions. In addition,
the great distance between the sensor and targets
means that some kinds of targets occupy only a few to
dozens of pixels in the imaging plane and are presented as
small objects that can easily be overwhelmed by a bright
background [6]. Objects of this kind are usually characterized
by a low signal-to-noise ratio (SNR) and inadequate
structure information, which is presented as a weak feature
response. These characteristics make the detection of small,
weak targets a more challenging task in remote sensing.
The past decade has witnessed major advances in object
detection in remote sensing images. At an early stage, various
models based on prior knowledge [7]-[10] were proposed
for target detection in satellite images. As image resolution
increases, prior-knowledge-based models increase in
uncertainty because the high complexity of HRRS images
tends to cause limited detection accuracy. More recently,
various forms of machine learning (ML) approaches [11]
have played a critical role in object detection. With the increasing
availability of big data and remarkable advances in
data mining, novel methods have come into use for HRRS
image processing.
Deep learning (DL) models [12]-[15] have attracted serious
attention and become dominant tools for processing
large-scale, high-dimension data; they have achieved satisfactory
accuracy for several tasks in the field of remote sensing.
By stacking multiple nonlinear layers, DL models extract
semantic information about objects as well as the context
relationships among them and the background. DL models
demonstrate superiority in the extractions and fusions
of multiscale features and have therefore outperformed the
early models, with significant developments in remote sensing
object representations. In recent decades, many works
have presented ML- and DL-based models, leading to the
creation of a series of benchmark data sets for promoting
remote sensing and small, weak object detection [16]-[19].
Although several survey papers on object detection have
been published, they have focused mainly on detection
technologies from the image-processing aspect [20], [21] or
on reviewing some categories of approaches, such as ML[11]
and DL-based methods [19], or some specific detection
problems and tasks, including vehicle detection [22] and salient
methods [23]-[25] for remote sensing object detection.
There is still the lack of a comprehensive review of existing
works that addresses the problems of small, weak object detection.
Based on the aforementioned analysis, this article
DECEMBER 2021 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
9
©SHUTTERSTOCK.COM/WILLEM
http://www.SHUTTERSTOCK.COM/WILLEM
IEEE Geoscience and Remote Sensing Magazine - December 2021

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