Syntax Literate: Jurnal Ilmiah Indonesia p–ISSN: 2541-0849 e-ISSN: 2548-1398
Vol.
9, No. 12, Desember 2024
CLOUD REMOVAL BASED ON
DARK CHANNEL PRIOR: A SYSTEMATIC LITERATURE REVIEW
Nazifa Hamidiyati1, Laksmita Rahadianti2
Universitas Indonesia, Indonesia1,2
Email:
[email protected]1, [email protected]2
Abstract
Remote sensing
satellite technology has revolutionized the way we gather information about our
planet. Through the use of advanced imaging capabilities, satellite images have
become invaluable in various aspects of daily life. These images are
extensively utilized in environmental protection, agricultural engineering, and
other fields. Remote sensing satellite maps are used for tasks such as
geological mapping, monitoring urban heat islands, environmental surveillance,
and detecting forest fires from remote sensing images. However, clouds present
a significant hindrance when utilizing satellite imagery for ground
observations, as they obstruct the view and can limit the accuracy of the
analysis. While there are numerous advanced state-of-the-art approaches
available, it is important to note that they often require a substantial amount
of data for training. On the other hand, if a more general approach is desired
without the need for extensive training data, pixel-based methods provide a
viable option. One of the widely used pixel-based methods for cloud removal in
satellite images is Dark Channel Prior (DCP). DCP is often combined with other
methods to improve the image quality. This systematic literature review will
demonstrate the development of the DCP method in cloud removal from satellite
images.
Keywords: Cloud
removal, dark channel prior (DCP), satellite imagery, remote sensing
Introduction
The Dark Channel Prior (DCP) theory, proposed by
The main principle behind Dark Channel Prior is derived from
haze-free outdoor scenes. In a haze-free image, the dark channel exhibits very
low intensity values that are close to zero, resulting in a dark appearance. By
normalizing the image, the pixels with the highest intensity, representing the
brightest 0.1% of pixels, are considered as skylight. It is assumed that the
color of atmospheric light closely resembles the color of the sky and is always
a positive value. The dark channel is defined as the minimum intensity across
the RGB channels, with at least one color channel
having very low values near zero
Monitoring the atmospheric light using the dark channel prior is a
reliable approach, especially when a large local mask is used to evaluate the
dark channel. Therefore, if the local mask size used for dark channel
evaluation is not large enough, it is recommended to utilize an additional dark
channel with a larger local mask size specifically for atmospheric light
monitoring. The use of local entropy is also found to be beneficial in
enhancing monitoring accuracy as it helps prevent the inclusion of intense
objects in atmospheric light estimation
In order to estimate the scattered light,
determine the lowest intensity within a patch for each HSI image containing
clouds. The intensity in the HSI color space reflects the level of brightness,
allowing the opacity of clouds to be inferred from this intensity. Through
observations, it has been noticed that cloud pixels often exhibit significantly
higher intensities compared to cloud-free pixels, which can be attributed to
the clouds appearing as white or grey sheets in the image
Based on the model of atmospheric scattering in
satellite imagery, the bright pixels within the satellite area are influenced
by the hazy atmospheric light. To estimate the atmospheric light, a
decomposition technique called Atmospheric Light Estimation (ALE) utilizes
local statistics to preserve image details. Unlike filters that tend to blur
out noise with higher variances, this adaptive filter effectively smooths out
noise with lower variation. As a result of its adaptable nature, this filter is
able to retain features in both low and high-contrast areas
Review Methods
To demonstrate the advancements in dark channel
detection methods for cloud removal in satellite imagery, we conducted a
systematic review of the existing literature. This was done to showcase the
success rate of previous methods and propose potential methods that could be
utilized in the future. The procedure consists of 6 steps, represented in the
diagram below.
Figure 1. The
Review Procedure
A. Formulation of Research Questions
The research questions were formulated to establish the
scope of discussion during the systematic literature review (SLR). The research
questions can be seen in the following Table 1.
Table 1. Research Questions
|
ID |
Research Questions |
|
RQ1 |
How is the DCP method approach utilized for cloud removal? |
|
RQ2 |
What are the proposed modifications to enhance the
performance of the DCP method for cloud removal? |
B. Identification of Database
The articles gathered and analyzed in this study are
sourced from a comprehensive scientific database, comprising six reputable
platforms: IEEE Xplore, Scopus, Science Direct, SpringerLink, and ACM Digital
Library.
C. Article Search Methodology
The keywords used for the search were: (‘cloud removal’ or
'cloud') and (‘dark channel prior’ or 'dark channel' or 'dark pixel'). The
following criteria, as presented in Table 2, were utilized to identify the candidate
articles.
Table 2.
Articles Criteria
|
Population |
Remote sensing image |
|
Intervention |
The approach of the DCP method |
|
Comparison |
Modification of the DCP method |
|
Outcome |
Improvement of DCP performance |
|
Context |
Cloud removal |
Using the specified keywords and criteria, we choose the
articles that were published in 2017-2022. The initial phase of the article
search yielded from the selected database can be seen in the following Table 3.
Table 3. Article
Search Result from Each Database
|
Database |
Number of Articles |
|
IEEE Xplore |
34 |
|
Scopus |
105 |
|
Science Direct |
8632 |
|
SpringerLink |
15639 |
|
ACM Digital Library |
2209 |
D. Article Selection
The initial phase of the article selection involves
applying inclusion and exclusion criteria, which are detailed in Table 4.
Table 4. Article
Inclusion and Exclusion Criteria
|
Inclusion |
The article is written in English |
|
The article is published between 2017-2022 |
|
|
The full text of the article can be accessed |
|
|
The article is related to the search keyword query |
|
|
The article can answer the research questions |
|
|
The article discusses the DCP method modification |
|
|
Exclusion |
The article is not written in English |
|
The article is not published between 2017-2022 |
|
|
The full text of the article can not
be accessed |
|
|
The article is not related to the search keyword query |
|
|
The article does not answer the research questions |
|
|
The article does not discuss the DCP method modification |
E. Doing the Review
Table 5 will provide the articles obtained from each
database after applying the inclusion and exclusion criteria. The systematic
literature review (SLR) will be conducted based on the selected articles.
Table 5.
Article Selection Result
|
Database |
Number of Articles |
|
IEEE Xplore |
13 |
|
Scopus |
22 |
|
Science Direct |
30 |
|
SpringerLink |
25 |
|
ACM Digital Library |
17 |
F. Synthesizing the Results
The purpose of this section is to address our research
questions and gather relevant information from the selected articles. The results
will be presented in the next section.
Results and Discussion
Based on the conducted literature review of
selected articles, it can be concluded that in traditional DCP methods,
modifications are made by combining DCP with specific techniques. On the other
hand, in learning-based methods, modifications are made to the model used to
determine the atmospheric light, as the atmospheric light utilizes the dark
channel prior as an approach for atmospheric light estimation. The summary of
the results can be seen in the following Table 6.
Table 6. Results
Summary
|
Ref |
Methods |
Description |
|
|
Modified
Dark Channel Prior with Multiple Scale (MDCPMS) |
·
For thin-cloud removal ·
Multiple scales and DCP are integrated together. ·
Multi-scale decomposition. ·
Decomposition of the input
image into high-frequency and low-frequency components. ·
The low-frequency component
is processed using DCP. ·
The resulting image
exhibits minimal color distortion. ·
The evaluation metrics
indicate superior results
compared to DCP and non-local methods. |
|
|
Dark channel subnet (NGAD) |
·
Based on the Gabor transform and Attention modules ·
Encoder–decoder structure and incorporated with Dark channel
subnet ·
The feature map is reconstructed by dark channel subnet with the Spatial attention module |
|
|
Sphere
Model Improved Dark Channel Prior |
·
The sphere model is
utilized to enhance DCP based on the radius (R) of the sphere
used to determine the minimum pixel intensity in cloud-contaminated areas. ·
The radius (R) approximates the standard deviation of pixels in the local patch. ·
The DCP algorithm is
subsequently employed after obtaining the transmission map based on the
Sphere model. ·
The process is
relatively slow. |
|
|
Deep convolutional autoencoder |
· Thin-cloud
removal ·
Utilizing
the simple linear iterative clustering (SLIC) superpixel segmentation method ·
Using a deep convolutional autoencoder for dehazing aerial images ·
Generating
a dehazed version of the image
without the need for additional
information like transmission map or atmospheric light value ·
Treating
dehazing as a one-step
problem rather than a two-step process to prevent error
amplification between stages · Utilizing the Adam optimization approach to minimize the
loss function. |
|
|
Low-Rank
and Sparse Constrained Dark Channel Prior |
·
Involving Lagrange multipliers and DCP. ·
Decomposing the scatter matrix
and low-rank matrix using Lagrange multipliers to remove thick
clouds. ·
Utilizing Dark Channel Prior (DCP) to remove thin
clouds. ·
The DCP method works
well in removing thin clouds but
cannot eliminate thick clouds. ·
The proposed method
can remove both thin and
thick clouds while reconstructing information from cloud-covered areas. |
|
|
Multiscale
Dark Channel Prior (MDCP) |
·
The problem of cloud
removal is transformed into image fusion. ·
The combination of
multiscale transformation
and sparse representation (SR) is used for traditional
image fusion. ·
DCP is integrated
into this fusion framework. ·
The fusion framework
applies the low-frequency component after the multiscale
transformation (MST) process.
·
Modified traditional Laplacian sharpening operations are employed to enhance
the sharpness of the results. |
|
|
Multimodel Fusion |
·
U-Net and SegNet
are utilized as the base network models. ·
The AdaBoost algorithm
is employed to enhance the
cloud feature extraction results. ·
Test-time augmentation (TTA) and the DCP model are used in the post-processing layer to improve accuracy
and edge extraction effects. ·
The DCP model can enhance
the similarity level of LFIs in high-resolution
satellite images to detect and
remove thin clouds. |
|
|
Robust gamma-correction-based
dehazing model (RGDM) |
· For thin-cloud removal · The Scene Albedo Restoration
Formulae (SARF) is employed to streamline the subsequent refinement process
and effectively handle the issue of non-uniform illumination · In the
proposed SARF, the patch size of
DCP, patch size of BCP, and mean
filter size need to be manually
initialized · The
traditional gamma correction technique is used to approximate the atmospheric
scattering model, which can be represented by an exponential form. |
|
|
Three-step post-processing strategy |
·
Based on radiation transmittance differences between cloud-covered and non-cloud-covered landscapes ·
The radiation transmittance is estimated based on the dark
channel prior (DCP) ·
The overestimated radiation
transmittance is corrected using spectral features ·
The modified radiation transmittance map for cloud detection is focused on reducing the impact of buildings
with high reflectance on cloud detection |
|
|
Temporal
information injection network (TIIN) |
· Parallel CNN-based architecture · a cirrus band from a cloudy region was employed as a reference to simulate
a nonuniform haze cover · The group convolution block consisting of three layers
is used for feature extension
by extracting multiscale semantic and contextual information. · only one channel
is considered consistently for the convolutional layers to achieve
spatial attention |
|
|
Sparse dark pixel region detection |
· Using the thin-cloud mask · A nonparametric measure is used to
evaluate the density of local
dark pixels · The
region with the sparse dark pixel
is selected as the thin-cloud candidate · The
multispectral images obtained by the
Wide Field View (WFV) |
|
|
Combination of the traditional
method and deep learning method |
·
U-Net is utilized to estimate
the reference thin cloud thickness
map from the original cloudy image. ·
The thin cloud thickness map for each spectral
band is obtained by searching for dark pixels
within a local window in the cloudy image. · A novel CNN architecture called
Slope-Net is developed to obtain the thin
cloud thickness map for each spectral
band. ·
A new method for simulating
wavelength-dependent thin clouds is proposed to generate suitable multispectral
cloudy images for training U-Net and Slope-Net. |
|
|
Wasserstein generative adversarial network (WGAN) in YUV
color space (YUV-GAN) |
· End-to-end thin cloud removal by learning luminance
and chroma components · the
generator adopts a residual
encoding–decoding network · adequate simulated pairs were used to train
the YUV-GAN · cloud-free images in RGB color space can be
obtained by the post-processing Inverse Transformation operation. |
|
|
The principal component pursuit
and alternating direction multiplier method (PCP-ADMM) |
·
Incorporating
the low-rank and sparse prior (LSP) concept · The dark channel of a
hazy image is decomposed into two components: the dark channel of direct
attenuation with sparsity and the atmospheric veil with low rank · The PCP-ADMM algorithm
is employed for low-rank and sparse decomposition to obtain an initial
estimation of the atmospheric veil · The refined atmospheric
veil is then utilized to estimate the atmospheric light |
|
|
Spectral grouping network (SG-Net) |
· For thin-cloud removal · Groups each HSI into
several spectral subsets based on the intra-spectral correlations · The
spectral correlation in
HSI is considered in the SG-Net-based data-driven method for
thin-cloud removal · Using different attention modules
to transfer useful information among the spectral bands |
Based on the literature review results, various methods
have been developed to address the issue of thin cloud removal in satellite and
aerial imagery. Traditional methods such as Dark Channel Prior (DCP) have been
modified with multi-scale approaches and mathematical model integration, such
as sphere models and frequency decomposition methods, to enhance accuracy in
thin cloud removal. On the other hand, machine learning and deep learning-based
methods, such as convolutional autoencoders and neural networks, provide more
efficient and accurate solutions by utilizing superpixel
segmentation and parallel convolution models. These methods overcome the
limitations of traditional DCP by reducing color distortion and improving the
accuracy of atmospheric light estimation, resulting in higher-quality images
without the need for separate steps in the dehazing process.
Furthermore, the integration of new techniques such as
Generative Adversarial Networks (GANs), U-Net convolutional networks, and
fusion models demonstrates significant potential in enhancing thin cloud
removal outcomes. These techniques enable more realistic cloud image
simulations and leverage spectral transformations to correct uncertainties in
transmittance estimation. Multimodel and fusion
approaches, such as combining U-Net and SegNet, offer
more precise results by sharpening edges and improving accuracy in cloud
removal. With these advancements, the dehazing process becomes more efficient,
providing a one-step solution that minimizes cumulative errors between stages,
supporting enhanced image quality for further applications in remote sensing
and mapping.
Conclusion
Based on the conducted review, several
conclusions can be drawn: (1) Modified DCP methods have shown improved results
in both cloud removal tasks and enhancing the color quality of images. (2) There
is an increasing trend in utilizing learning-based approaches for modifying DCP
methods. (3) The majority of modifications to DCP methods have been focused on
thin-cloud removal, while modifications specifically targeting thick-cloud
removal are still limited.
BIBLIOGRAPHY
Anwar, M. I., & Khosla, A.
(2017). Vision enhancement through single image fog removal. Engineering
Science and Technology, an International Journal, 1075-1083.
Bi, G., Si, G., Zhao, Y., Qi, B., &
Lv, H. (2022). Haze Removal for a Single Remote Sensing Image Using Low-Rank
and Sparse Prior. IEEE Transactions on Geoscience and Remote Sensing.
Cheng, J., Zhang, Y., Zhou, X., &
Shi, S. (2021). A Low-Rank and Sparse Constrained Dark Channel Prior for
Cloud Removal in Remote Sensing Image Sequence. 2021 IEEE International
Geoscience and Remote Sensing Symposium IGARSS.
Fazlali, H., Shirani, S., McDonald, M.,
Brown, D., & Kirubarajan, T. (2020). Aerial image dehazing using a deep
convolutional autoencoder. Multimedia Tools and Applications.
Gao, J., Yuan, Q., Li, J., & Su, X.
(2021). Unsupervised missing information reconstruction for single remote
sensing image with Deep Code Regression. International Journal of Applied
Earth Observation and Geoinformation, 102599.
He, K., Sun, J., & Tang, X. (2009).
Single image haze removal using dark channel prior. 2009 IEEE Conference
on Computer Vision and Pattern Recognition.
Ju, M., Ding, C., Guo, Y. J., &
Zhang, D. (2018). Remote Sensing Image Haze Removal Using
Gamma-Correction-Based Dehazing Model. Digital Object Identifier.
Kumar, N. U., Shivakumar, N., &
Bachu, S. (2022). Satellite Image Dehazing Using Fast Iterative Domain
Gaussian Guided Image Filtering. 2022 3rd International Conference on
Smart Electronics and Communication (ICOSEC), (pp. 1430-1435).
Li, J., Hu, Q., & Ai, M. (2019). Haze
and Thin Cloud Removal via Sphere Model Improved Dark Channel Prior. IEEE
Geoscience and Remote Sensing Letters, 472-476.
Lin, Y., He, L., Zhang, Y., & Wu, Z.
(2022). Cloud Detection of Gaofen-2 Multi-Spectral Imagery Based on the
Modified Radiation Transmittance Map. Remote Sensing, 14.
Liu, F., Lv, Y., Li, B., Gao, S., &
Qin, Y. (2021). A Semiphysical Approach of Haze Removal for Landsat Image. Journal
of Selected Topics in Applied Earth Observations and Remote Sensing,
7410-7421.
Ma, X., Wang, Q., & Tong, X. (2022).
A spectral grouping-based deep learning model for haze removal of
hyperspectral images. ISPRS Journal of Photogrammetry and Remote Sensing,
177-189.
Ma, X., Wang, Q., Tong, X., &
Atkinson, P. M. (2022). A deep learning model for incorporating temporal
information in haze removal. Remote Sensing of Environment, 113012.
Markchom, T., & Lipikorn, R. (2018).
Thin Cloud Removal Using Local Minimization and Logarithm Image
Transformation in HSI Color Space. 2018 4th International Conference on
Frontiers of Signal Processing (ICFSP).
Shi, S., Zhang, Y., Zhou, X., &
Cheng, J. (2021). Cloud Removal for Single Visible Image Based on Modified
Dark Channel Prior with Multiple Scale. 2021 IEEE International Geoscience
and Remote Sensing Symposium IGARSS, (pp. 4127-4130).
Shi, S., Zhang, Y., Zhou, X., &
Cheng, J. (2022). A Novel Thin Cloud Removal Method Based on Multiscale Dark
Channel Prior (MDCP). IEEE Geoscience and Remote Sensing Letters, 1-5.
Singh, D., & Kumar, V. (2018).
Comprehensive survey on haze removal techniques. Multimedia Tools and
Applications.
Tang, X., Yao, J., Chen, J., Li, G.,
& Zhang, W. (2022). Multimodel Fusion Method for Cloud Detection in
Satellite Laser Footprint Images. IEEE Geoscience and Remote Sensing
Letters, 1-5.
Wen, X., Pan, Z., Hu, Y., & Liu, J.
(2021). Generative Adversarial Learning in YUV Color Space for Thin Cloud
Removal on Satellite Imagery. Remote Sensing, 13.
Wu, W., Luo, J., Hu, X., Yang, H., &
Yang, Y. (2018). A Thin-Cloud Mask Method for Remote Sensing Images Based on
Sparse Dark Pixel Region Detection. Remote Sensing.
Zhang, C., Zhang, X., Yu, Q., & Ma,
C. (2022). An Improved Method for Removal of Thin Clouds in Remote Sensing
Images by Generative Adversarial Network. IGARSS 2022 - 2022 IEEE
International Geoscience and Remote Sensing Symposium, (pp. 6706-6709).
Zhang, J., Wang, X., Yang, C., Zhang, J.,
He, D., & Song, H. (2018). Image dehazing based on dark channel prior and
brightness enhancement for agricultural remote sensing images from
consumer-grade cameras. Computers and Electronics in Agriculture,
196-206.
Zhang, J., Zhou, Q., Wu, J., Wang, Y.,
Wang, H., Li, Y., . . . Liu, Y. (2020). A Cloud Detection Method Using
Convolutional Neural Network Based on Gabor Transform and Attention Mechanism
with Dark Channel Subnet for Remote Sensing Image. Remote Sensing, 12.
Zi, Y., Xie, F., Zhang, N., Jiang, Z.,
Zhu, W., & Zhang, H. (2021). Thin Cloud Removal for Multispectral Remote
Sensing Images Using Convolutional Neural Networks Combined With an Imaging
Model. Journal of Selected Topics in Applied Earth Observations and Remote
Sensing, 3811-3823.
|
Copyright holder: Nazifa Hamidiyati, Laksmita Rahadianti (2024) |
|
First publication right: Syntax Literate: Jurnal Ilmiah Indonesia |
|
This article is licensed under:
|
