IEEE Geoscience and Remote Sensing Magazine - December 2023 - 29

framework that requires only image-level labels to identify
wildlife.
FOREST ECOSYSTEM PROTECTION
The forest ecosystem plays an important role in ecological
protection, climate regulation, and carbon cycling.
Understanding the condition of trees is essential for forest
ecosystem protection [324], [325], [326], [327], [328].
Safonova et al. [326] analyzed the shape, texture, and
color of detected trees' crowns to determine their damage
stage, providing a more efficient way to assess forest
health. Sani-Mohammed et al. [328] utilized an
instance segmentation approach to map standing dead
trees, which is imperative for forest ecosystem management
and protection.
FUTURE DIRECTIONS
Apart from the five RSOD research topics mentioned in
this survey, there is still much work to be done in this field.
Therefore, we present a forward-looking discussion of future
directions to further improve and enhance the detectors
in remote sensing scenes.
UNIFIED DETECTION FRAMEWORK FOR
LARGE-SCALE REMOTE SENSING IMAGES
Benefiting from the development of remote sensing technology,
high-resolution large-scale RSIs (e.g., more than
10 ,,
000 10 000#
pixels) can be easily obtained. However,
limited by GPU memory, the current mainstream RSOD
methods fail to directly perform object detection in largescale
RSIs but adopt a sliding window strategy, mainly including
sliding window cropping, patch prediction, and
results merging. On the one hand, this sliding window
framework requires complex data preprocessing and postprocessing,
compared with the unified detection framework.
On the other hand, objects usually occupy a small
area of RSIs, and the invalid calculation of massive backgrounds
leads to increasing computation time and memory
consumption. Some studies [215], [329], [330] proposed a
coarse-to-fine detection framework for object detection in
large-scale RSIs. This framework first locates ROIs by filtering
out meaningless regions and then achieves accurate detection
from these filtered regions.
DETECTION WITH MULTIMODAL
REMOTE SENSING IMAGES
Restricted by the sensor imaging mechanism, detectors
based on single-modal RSIs often have detection performance
deviations, which are difficult to meet in practical
applications [331]. In contrast, multimodal RSIs from different
sensors have their own characteristics. For instance,
hyperspectral
images contain high-spectral-resolution
and fine-grained spectral features, SAR images provide
abundant texture information, and optical images exhibit
high spatial resolution with rich detailed information. The
integrated processing of multimodal RSIs can improve the
DECEMBER 2023 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
interpretation of scenes and obtain a more objective and
comprehensive understanding of geospatial objects [332],
[333], [334], providing the possibility to further improve
the detection performance of RSOD.
DOMAIN ADAPTATION OBJECT DETECTION
IN REMOTE SENSING IMAGES
Due to the diversity of remote sensing satellite sensors,
resolutions, and bands, as well as the influence of weather
conditions, seasons, and geospatial regions [6], RSIs collected
from different satellites are generally drawn from
similar but not identical distributions. Such distribution
differences (also called the domain gap) severely restrict the
generalization performance of the detector. Recent studies
on domain adaptation object detection [335], [336], [337],
[338] have proposed to tackle the domain gap problem.
However, these studies focus only on domain adaptation
detectors in single-modal RSIs, while cross-modal domain
adaptation object detection (e.g., from optical images to
SAR images [339], [340]) is a more challenging and worthwhile
topic to investigate.
INCREMENTAL DETECTION OF
REMOTE SENSING OBJECTS
The real-world environment is dynamic and open, where
the number of categories has evolved over time. However,
mainstream detectors require both old and new data to
retrain the model when meeting new categories, resulting
in high computational costs. Recently, incremental learning
has been considered the most promising way to solve
this problem, as it can learn new knowledge without forgetting
old knowledge while using only new data [341]. Incremental
learning has been preliminarily explored in the
remote sensing community [342], [343], [344], [345]. For
example, Chen et al. [342] integrated knowledge distillation
into FPN and detection heads to learn new concepts
while maintaining old ones. More thorough research is still
needed in incremental RSOD to meet the dynamic learning
task in practical applications.
SELF-SUPERVISED PRETRAINED MODELS
FOR REMOTE SENSING SCENES
Current RSOD methods are always initialized with ImageNet
[346] pretrained weights. However, there is an inevitable
domain gap between natural and remote sensing
scenes, probably limiting the performance of RSOD.
Recently, self-supervised pretraining approaches have
received extensive attention and shown excellent performance
in classification and downstream tasks in nature
scenes. Benefiting from rapid advances in remote
sensing technology, abundant remote sensing data
[347], [348] also provide sufficient data support for selfsupervised
pretraining. Some researchers [349], [350],
[351], [352], [353] have initially demonstrated the effectiveness
of remote sensing pretraining on representative
downstream tasks. Therefore, exploring self-supervised
29
IEEE Geoscience and Remote Sensing Magazine - December 2023

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - December 2023
