IEEE Geoscience and Remote Sensing Magazine - December 2021 - 27

and how to transfer well-trained detectors to different modalities
of data.
WEAKLY SUPERVISED DETECTION
Recent state-of-the-art approaches require many samples
with accurate annotation, in the manner of fully supervised
learning. However, labeling samples is labor intensive and
time consuming. Meanwhile, weakly/partially annotated
or unlabeled samples are easily accessible and sufficient.
Therefore, it is essential to leverage DL-based models to
learn from these samples to boost detection ability.
LIGHTWEIGHT OBJECT MODEL
The number of layers in existing CNN models for extracting
features has dramatically increased from several [14] to hundreds
of layers [13], [206]. They have millions of parameters
and need massive computation resources and training data
to obtain an ideal mode. To train the CNN models effectively,
much work has been done to develop a series of lightweight
and compact models. However, a significant gap in the efficiency
between detectors and the human eye remains.
AUTOMATIC NEURAL ARCHITECTURE SEARCH
Most existing target detectors are based on manual design.
To meet problems of ever-increasing complexity requires increasing
domain knowledge and expertise. Recently, a natural
research direction has been to automatically select and build
a detector with a performance that can deal with the number
of parameters, such as automated ML [201]. Related work
should be carried out for small, weak object detection.
IMPROVEMENT OF IMAGE QUALITY
Affected by imaging conditions such as weather, light, and
the resolution of sensors, remote sensing images may not be
able to meet the requirements of usage, as they are blurred
or noisy or have low resolution. Algorithms, such as those
undertaken in image fusion, image denoising, and superresolution,
have been developed to address these problems.
These should be combined with detection methods to improve
detection performance.
UNIVERSAL OBJECT FRAMEWORK
Recently, increasing efforts have been made in learning universal
representations, reinforcement learning, and lifelong
learning; these are effective in learning, transferring, and
reasoning knowledge from massive data. It is meaningful to
design a universal object framework based on state-of-theart
advances, which can gradually self-evolve and improve
detection performance.
CONCLUSIONS
To meet the requirements of some applications, the task of
small, weak object detection, which is more challenging
than generic object detection, has gradually become increasingly
important and attracted much attention. During
the last several years, considerable efforts have been made
DECEMBER 2021 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
to develop various methods that address small, weak object
detection. This article presented a systematic review of
the advances of small, weak object detection in the remote
sensing community. Having analyzed the challenges and
difficulties of small, weak target detection, we discussed
the technical evolution of object detection and benchmark
data sets. Finally, we categorized the existing works that address
different challenges and in which some promising research
directions have been drawn for the further improvement
of small, weak object detection. The research of small,
weak object detection is still far from complete, but given
the breakthroughs over the past several years, we are optimistic
about future developments.
ACKNOWLEDGMENTS
This work was supported by the National Natural Science Foundation
of China under grants U1711266 and 41925007 and
the Fundamental Research Funds for the Central Universities,
China University of Geosciences, Wuhan (no. 162301212697).
Lizhe Wang and Ruyi Feng are the corresponding authors.
AUTHOR INFORMATION
Wei Han (weihan@cug.edu.cn) is with the School of Computer
Science, China University of Geosciences, Wuhan,
430078, China, and also the Hubei Key Laboratory of Intelligent
Geo-Information Processing, China University of
Geosciences, Wuhan, 430078, China.
Jia Chen (chen_jia@cug.edu.cn) is with the School of
Computer Science, China University of Geosciences, Wuhan,
430078, China, and also the Hubei Key Laboratory of
Intelligent Geo-Information Processing, China University
of Geosciences, Wuhan, 430078, China.
Lizhe Wang (lizhe.wang@foxmail.com) is with the
School of Computer Science, China University of Geosciences,
Wuhan, 430078, China, the Hubei Key Laboratory
of Intelligent Geo-Information Processing, China University
of Geosciences, Wuhan, 430078, China, and the Key
Laboratory of Geological Survey and Evaluation of the
Ministry of Education, China University of Geosciences,
Wuhan, 430078, China.
Ruyi Feng (fengry@cug.edu.cn) is with the School of
Computer Science, China University of Geosciences, Wuhan,
430078, China, and also the Hubei Key Laboratory of
Intelligent Geo-Information Processing, China University
of Geosciences, Wuhan, 430078, China.
Fengpeng Li (li_feng_peng@cug.edu.cn) is with the
School of Computer Science, China University of Geosciences,
Wuhan, 430078, China, and also the Hubei Key Laboratory
of Intelligent Geo-Information Processing, China
University of Geosciences, Wuhan, 430078, China.
Lin Wu (wulin@cug.edu.cn) is with the Key Laboratory of
Geological Survey and Evaluation of the Ministry of Education,
China University of Geosciences, Wuhan, 430074, China.
Tian Tian (tiantian@cug.edu.cn) is with the School of
Computer Science, China University of Geosciences, Wuhan,
430078, China, and also the Hubei Key Laboratory of
27
IEEE Geoscience and Remote Sensing Magazine - December 2021

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