Next Article in Journal
Environment Mapping-Based Classification for Reverse Engineering Using Supervised Learning
Next Article in Special Issue
KARAN: Mitigating Feature Heterogeneity and Noise for Efficient and Accurate Multimodal Medical Image Segmentation
Previous Article in Journal
Task-Level Customized Pruning for Image Classification on Edge Devices
Previous Article in Special Issue
DDE-Net: Dynamic Density-Driven Estimation for Arbitrary-Oriented Object Detection
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Electronics
Volume 13
Issue 20
10.3390/electronics13204030
electronics-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Atmospheric Gravity Wave Detection in Low-Light Images: A Transfer Learning Approach

by
Beimin Xiao
1,
Shensen Hu
1,2,3,*,
Weihua Ai
1,3 and
Yi Li
1
1
College of Meteorology and Oceanography, National University of Defense Technology, Changsha 410073, China
2
School of Atmospheric Sciences, Nanjing University, Nanjing 210093, China
3
Key Laboratory of High Impact Weather (Special), China Meteorological Administration, Changsha 410073, China
*
Author to whom correspondence should be addressed.
Electronics 2024, 13(20), 4030; https://doi.org/10.3390/electronics13204030
Submission received: 15 June 2024 / Revised: 8 October 2024 / Accepted: 11 October 2024 / Published: 13 October 2024
(This article belongs to the Special Issue Artificial Intelligence in Image and Video Processing)
Browse Figures
Versions Notes

Abstract

:
Atmospheric gravity waves, as a key fluctuation in the atmosphere, have a significant impact on climate change and weather processes. Traditional observation methods rely on manually identifying and analyzing gravity wave stripe features from satellite images, resulting in a limited number of gravity wave events for parameter analysis and excitation mechanism studies, which restricts further related research. In this study, we focus on the gravity wave events in the South China Sea region and utilize a one-year low-light satellite dataset processed with wavelet transform noise reduction and light pixel replacement. Furthermore, transfer learning is employed to adapt the Inception V3 model to the classification task of a small-sample dataset, performing the automatic identification of gravity waves in low-light images. By employing sliding window cutting and data enhancement techniques, we further expand the dataset and enhance the generalization ability of the model. We compare the results of transfer learning detection based on the Inception V3 model with the YOLO v10 model, showing that the results of the Inception V3 model are greatly superior to those of the YOLO v10 model. The accuracy on the test dataset is 88.2%.

1. Introduction

Atmospheric gravity waves, which are a perturbation of the density structure of the atmosphere, are one of the most significant fluctuations in the atmosphere. They are also a key process in the dynamics and thermodynamics of the atmosphere. Gravity waves transport energy and momentum to the stratosphere, mesosphere, and even higher altitudes through propagation in the vertical direction, continuously affecting the dynamics and physics of the atmosphere from turbulence to the atmospheric circulation, which has significant impacts on the structure and stability of the atmospheric circulation and plays an important role in both climate change and the development of the weather [1,2,3].
In the past decades, compared to the troposphere and stratosphere, there have been few gravity wave observations in the mesopause (~90 km altitude). The airglow optical imaging technique is currently the key tool for observing gravity waves in the middle and upper atmosphere. Atmospheric gravity waves are not directly optically observable, but they disturb the airglow emission layer during propagation, thus producing observable fluctuations in specific emission bands [4,5,6,7,8]. By monitoring the airglow in different spectral bands, it is possible to observe the perturbation phenomenon of gravity waves in the airglow layer at different altitudes. Currently, two main methods are employed for atmospheric gravity wave detection: The ground-based airglow imager and satellite-based low-light detectors.
The first airglow image of gravity waves was observed by all-sky airglow imagers (ASAIs) in 1982 by Peterson and Adams [9], and since then, ASAIs have been widely used for mesopause gravity wave observation and parameter extraction [5,10,11,12]. The China Meridian Space Weather Monitoring Project has established an airglow observation network since 2009 with the objective of accumulating more observation data [13]. Lai et al. [14] developed an automatic extraction procedure based on a convolutional neural network (CNN) and an object detection model based on faster region-based convolutional neural network (Faster R-CNN) to identify airglow gravity wave images and calculate their parameters. However, the observation of ground-based airglow imagers is not only affected by adverse weather conditions but also presents significant challenges in achieving a global scale, particularly in distant sea regions, due to the limitations of ground station deployment.
Satellite remote sensing has a wide observation range and is capable of obtaining gravity wave features from hundreds to even thousands of kilometers globally, which compensates for the shortcomings of airglow imagers. The S-NPP (Suomi NPP) satellite, launched by the United States in 2011, carries the Visible Infrared Imaging Radiometer Suite (VIIRS)/Day–Night Band (DNB), a low-light payload with powerful night-time detection capability. The low-light imager, which is designed to detect the atmosphere and ground environment under low-illumination conditions at night, was discovered to be capable of observing mesopause gravity waves during weak moonlight. Miller et al. [15] found for the first time that the presence of mesopause gravity waves in low-light images of S-NPP satellites often results in airglow radiation exhibiting a distinctive striped image feature. This provides a novel method of observing mesopause gravity waves at night. In contrast to ground-based optical measurements, satellite-based observations of gravity waves are not constrained by clouds and adverse weather conditions [16]. Due to the fact that polar-orbiting satellites transit only once a night and it is difficult to manually classify observation data accumulated over many years, existing studies on the excitation mechanism and parameters of mesopause gravity waves are inadequate.
In recent years, deep learning has demonstrated considerable potential in a range of disciplines and is widely employed in areas such as speech and image recognition [17]. CNNs represent a prominent example of deep learning algorithms and demonstrate efficacy in classification tasks. The seminal work of Hubel and Wiesel [18] established the potential of CNNs for image processing. In the field of machine learning, LeCun et al. [19] established the modern CNN structure by developing the LeNet-5 multilayer neural network for handwritten digit classification. This established the modern structure of CNNs. In comparison to traditional artificial neural networks, CNNs are capable of attaining a more profound internal structure with a reduced number of parameters, which significantly minimizes the quantity of model parameters and enhances the efficiency of the training process [20]. The design of a novel CNN structure for diverse datasets is anticipated to yield optimal performance. However, the process is time-consuming, and to enhance efficiency, successful networks are often integrated to align with the requisite specifications [21]. Examples of such networks include AlexNet, GoogleNet, VGGNet, XceptionNet, and others [22,23,24,25]. CNN models are capable of extracting features such as color, texture, and shape from an image, which is advantageous in satellite image processing and is commonly employed in tasks such as satellite image target identification. Kavita et al. [26] demonstrated that CNN performs well in classification tasks with small sample datasets. Wang et al. [27] employed the Inception V3 model of CNN to classify the atmospheric and oceanic phenomena contained in the WV model data of the Sentinel-1 satellite. The overall accuracy rate for ten atmospheric and oceanic phenomena exceeded 93%. Guo et al. [28] employed CNN models and transfer learning method to identify the wind stripe phenomenon from Gaofen-3 images with an accuracy of 93% and up to 96% after dataset enhancement. Transfer learning provides a route to identify gravity wave images from small-sample low-light datasets. However, there have been few attempts to automatically detect gravity waves. In addition to Lai et al.’s attempt to extract gravity waves from airglow imager images [14], Matsuoka et al. [29] used a deep neural network model to estimate the gravity wave parameters from reanalysis data. Sreekanth et al. [30] proposed incorporating kernels into Deep Dictionary Learning to increase the effectiveness of the detection of Wave Breaking events. González et al. [31] explored various preprocessing and transfer learning methods to detect gravity waves using low-light satellite datasets.
In this paper, we focus on gravity wave events in the South China Sea region and utilize a one-year low-light satellite dataset processed with wavelet transform noise reduction and light pixel replacement. The Inception V3 model is selected to carry out transfer learning and the results show better performance compared with the YOLO v10 model. The detection accuracy of the Inception V3 model in the test dataset reaches 88.2%. This shows that the gravity wave detection model obtained by transfer learning is able to identify the presence of gravity waves in low-light images in a better way, which provides further support to the collection of more gravity wave events in the future to study various parameters and excitation mechanisms.

2. Materials and Methods

2.1. Low-Light Data

The low-light imager VIIRS carried by the S-NPP and JPSS-1 satellites inherits the advantages of previous instruments such as MODIS (Moderate-resolution Imaging Spectroradiometer) and OLS (Operational Linescan System), and it is currently the imaging radiometer with the strongest integrated detection capability [32,33]. The DNB channel on the VIIRS sensor is used as a panchromatic solar reflectance band, with a spectral response range of 0.5–0.9 μm, a mowing width of 3000 km, and a constant spatial resolution of 750 m. The wide dynamic response range enables it to detect faint radiation with a radiance of 10−9–10−11 Wcm−2sr−1 [34,35].
The SDR dataset used in this paper is from the Comprehensive Large Array-data Stewardship System (CLASS) of the National Oceanic and Atmospheric Administration (NOAA) of the United States, which uses DNB channel data from the S-NPP and JPSS-1 satellites. The SVDNB product records the radiance data observed by the sensors after in-orbit radiometric calibration. Each VIIRS low-light image is 3072 pixels × 4064 pixels, with a spatial resolution of about 750 m. As a daily update, the SVDNB product operates in real time and is efficient, and it is able to provide low-light data for at least one night transit per day. We selected the VIIRS low-light observation data from two satellites in the South China Sea region for the whole year of 2020 for deep learning network training.

2.2. Deep Learning Model

Inception Net is one of the representative algorithms of the convolutional neural network (CNN). It was initially deployed by Google in the 2014 ImageNet large-scale visual recognition competition. In this paper, we utilize the Inception V3 model, which optimizes the structure of the Inception Module by splitting a larger 2D convolutional kernel into two smaller 1D convolutional kernels. The Inception V3 model reduces the number of parameters, which improves the ability of the model to suppress overfitting and reduces the top-5 error rate from 4.8% to 3.5%. Following pre-training on the ImageNet dataset, the Inception V3 model is capable of extracting general features, such as texture, from images, and it can be employed for the classification of satellite images through transfer training, which is suitable for the limited availability of data in this paper.
The Inception V3 model, with a 48-layer network structure and over 23 million trainable parameters, requires a large dataset, extensive time, and high hardware specifications to train a new model. In the absence of a substantial dataset, transfer training can be utilized to achieve recognition capabilities while avoiding the process of training all parameters of the model. As shown in Figure 1, the Inception V3 model typically consists of two main components: the part responsible for extracting image features and the part that uses the Softmax layer and fully connected layers for image classification. By leveraging the feature extraction part of the pre-trained Inception V3 model on the ImageNet dataset to extract features from low-light images, and by training the newly constructed Softmax layer and fully connected layers on the TensorFlow deep learning framework, the detection of gravity waves can be accomplished. The advantage of transfer training is that the majority of the Inception V3 structure, with its retained parameters, has already developed the ability to recognize image edges and textures. Training the newly added structures with a limited amount of data is sufficient to enable the new model to identify gravity waves within low-light images.

3. Data Preprocessing

This paper is based on the principles of transfer learning, utilizing VIIRS/DNB low-light images to train the Inception V3 model, with the objective of developing a model suitable for gravity wave detection. The principal process is depicted in Figure 2. Initially, the low-light images are preprocessed, and each scene is then divided from top to bottom and left to right into several 300 pixels × 300 pixels sub-images, which are sequentially numbered. Subsequently, the sub-images are manually classified into two categories: those that exhibit gravity waves and those that do not. They are then input into the Inception V3 model for training. Finally, the performance of the model is analyzed using a reserved validation set.
Initially, preprocessing is performed on the low-light images. Under the condition of the new moon phase, the upward radiation from the moon reflected by the ground is extremely weak. The primary natural light source contributing to the low-light images is the airglow radiation, and gravity waves can be observed under such illumination conditions. Firstly, more than 1400 data samples were obtained from two satellites, S-NPP and JPSS-1. The samples were produced in the South China Sea region in 2020. Samples which satisfied the conditions of the new moon were selected. Subsequently, we manually identified the presence of gravity waves in the low-light images. Finally, we selected 68 low-light images that contained gravity waves.
The direct use of low-light images for training the Inception V3 model may affect training performance due to the large original size of the images. Therefore, we segmented the 68 preprocessed complete images into several 300 pixels × 300 pixels sub-images utilizing a sliding window technique with a size of 150 pixels, resulting in a total of 33,592 sub-images. Through manual discrimination, we further selected images containing the striped features of gravity waves from the sub-images. After the selection process, it was found that out of the 33,592 sub-images, 1434 contained gravity waves.
Although we expanded the dataset by segmenting it with a sliding window, the number of sub-images containing gravity waves was relatively small. To enhance the generalization ability of the model and prevent overfitting during training, we employed three data augmentation techniques, flipping, rotating and enhancing contrast, which expanded the wave dataset to 5736. The dataset was divided into two classes: ‘wave’ for gravity waves and ‘nowave’ for no gravity waves. In order to keep the dataset balanced, 5736 images were kept in each class, resulting in a total of 11,472 images.
Furthermore, the low signal-to-noise ratio (SNR) of the low-light observation data, caused by the weak airglow radiation, may have an impact on the extraction of gravity wave features. To improve the quality of the data, we performed a wavelet transform noise reduction on the dataset [36]. Similarly, overly bright city lights and lightning pixels can affect the detection of gravity waves, and we made substitutions for pixels larger than 5 nWcm−2sr−1 using the light threshold method [37].
The preprocessed dataset was randomly divided into three parts: the training set, the validation set, and the test set. The ratio of the number of images in each part was 7:2:1.

4. Results

A 64-bit Core i9 CPU with Windows 11 was used for all testing. TensorFlow 2.16 has been installed on the PC via Anaconda and Python 3.10. In the training process, the Inception V3 model were trained for 700 epochs with a learning rate of 0.001 and a batch of 300.
Figure 3 illustrates the process of training the Inception V3 model. The green line represents the training set accuracy, and the yellow line represents the validation set accuracy. As shown in the figure, initially, the accuracy of both datasets increases rapidly with time, and after approximately 500 epochs, the change in accuracy tends to stabilize.
Table 1 presents the statistical results of the model training accuracy. The accuracy of the training set is about 89.7%, the accuracy of the validation set is about 86.9%, and the precision of the test set is 88.2%. The accuracy indicates that the Inception V3 model based on transfer learning can identify gravity waves in low-light images well.
Figure 4 illustrates the results of the YOLO v10 model. From the trends depicted in the figures, it is evident that as the training progressed, there was a consistent reduction in box loss, classification loss, and distribution focal loss for both the training and validation phases. These reductions were found to correlate with improved accuracy in bounding box prediction, object classification, and confidence in prediction, respectively. Precision and recall also show a general upward trend, which indicates that the ability of the model to detect gravity waves continued to improve. However, the final detection ability of the model is inferior.
We used the results of the best model obtained from YOLO v10 training for the detection of the testing set, and the results are shown in Table 2. It can be seen that the results obtained from transfer learning using the Inception V3 model are significantly better than those obtained using the YOLO v10 model.

5. Discussion and Conclusions

In this paper, the Inception V3 model based on transfer learning is applied to low-light images to explore the automatic detection of atmospheric gravity waves at night. Based on VIIRS/DNB data from two satellites, S-NPP and JPSS-1, obtained in the South China Sea region in 2020, the images are firstly classified into two classes of gravity waves and no gravity waves using manual discrimination to establish training and validation datasets. Subsequently, the Inception V3 model is trained with the application of transfer learning and data enhancement techniques, resulting in a detection accuracy of 88.2% for the test set. In comparison with the YOLO v10 model, we find that the Inception V3 model significantly outperforms the YOLO v10 model in the small-sample low-light dataset.
In the future, we will collect more low-light images on a global scale and over a longer time frame in order to further optimize the Inception V3 model. This will improve the detection accuracy and generalizability of the model, as well as providing assistance for subsequent research on gravity wave parameters and excitation mechanisms.

Author Contributions

Conceptualization, S.H.; Methodology, B.X. and S.H.; Software, B.X.; Validation, S.H.; Formal analysis, S.H.; Investigation, W.A.; Resources, W.A.; Data curation, W.A.; Writing—original draft, B.X.; Writing—review & editing, B.X.; Supervision, Y.L.; Project administration, Y.L.; Funding acquisition, Y.L. All authors have read and agreed to the published version of the manuscript.

Funding

This work was supported by the National Natural Science Foundation of China under grant numbers 42305150 and 42475016; the Hunan Provincial Natural Science Foundation of China under grant number 2023JJ40665; and the Independent Innovation Science Fund of the National University of Defense Technology under grant nos. 22-ZZCX-081.

Data Availability Statement

The data used in this article can be found at the following URL: https://www.aev.class.noaa.gov/saa/products/welcome;jsessionid=5F48696BB108474316362922C58C57BF, accessed on 4 March 2024. The Google Inception V3 used in this article can be found at the following URL: http://download.tensorflow.org/models/image/imagenet/inception-2015-12-05.tgz, accessed on 15 March 2024.

Acknowledgments

We thank NOAA for making satellite light products publicly available. We would like to express our gratitude to Xiang Shao for his significant contributions to the application of the YOLO v10 model. His input was instrumental in the development of the research.

Conflicts of Interest

The authors declare no conflicts of interest.

References

  1. Holton, J.R. The Influence of Gravity Wave Breaking on the General Circulation of the Middle Atmosphere. J. Atmos. Sci. 1983, 40, 2497–2507. [Google Scholar] [CrossRef]
  2. Fritts, D.C.; Alexander, M.J. Gravity wave dynamics and effects in the middle atmosphere. Rev. Geophys. 2003, 41, 1–64. [Google Scholar] [CrossRef]
  3. Alexander, M.J.; Barnet, C. Using satellite observations to constrain parameterizations of gravity wave effects for global models. J. Atmos. Sci. 2007, 64, 1652–1665. [Google Scholar] [CrossRef]
  4. Vadas, S.L.; Yue, J.; She, C.Y.; Stamus, P.A.; Liu, A.Z. A model study of the effects of winds on concentric rings of gravity waves from a convective plume near Fort Collins on 11 May 2004. J. Geophys. Res. Atmos. 2009, 114, 2156–2202. [Google Scholar] [CrossRef]
  5. Yue, J.; Vadas, S.L.; She, C.Y.; Nakamura, T.; Reising, S.C.; Liu, H.L.; Stamus, P.; Krueger, D.A.; Lyons, W.; Li, T. Concentric gravity waves in the mesosphere generated by deep convective plumes in the lower atmosphere near Fort Collins, Colorado. J. Geophys. Res. Atmos. 2009, 114, D06104. [Google Scholar] [CrossRef]
  6. Miller, S.D.; Straka, W.C.; Yue, J.; Smith, S.M.; Alexander, M.J.; Hoffmann, L.; Setvák, M.; Partain, P.T. Upper atmospheric gravity wave details revealed in nightglow satellite imagery. Proc. Natl. Acad. Sci. USA 2015, 112, E6728–E6735. [Google Scholar] [CrossRef]
  7. Azeem, I.; Yue, J.; Hoffmann, L.; Miller, S.D.; Straka, W.C.; Crowley, G. Multisensor profiling of a concentric gravity wave event propagating from the troposphere to the ionosphere. Geophys. Res. Lett. 2015, 42, 7874–7880. [Google Scholar] [CrossRef]
  8. Azeem, I.; Barlage, M. Atmosphere-ionosphere coupling from convectively generated gravity waves. Adv. Space Res. 2017, 61, 1931–1941. [Google Scholar] [CrossRef]
  9. Peterson, A.W.; Adams, G.W. OH airglow phenomena during the 5–6 July 1982 total lunar eclipse. Appl. Opt. 1983, 22, 2682–2685. [Google Scholar] [CrossRef]
  10. Xu, J.; Li, Q.; Yue, J.; Hoffmann, L.; Straka, W.C.; Wang, C.; Liu, M.; Yuan, W.; Han, S.; Miller, S.D. Concentric gravity waves over northern China observed by an airglow imager network and satellites. J. Geophys. Res. Atmos. 2015, 120, 11058–11078. [Google Scholar] [CrossRef]
  11. Suzuki, S.; Shiokawa, K.; Otsuka, Y.; Ogawa, T.; Nakamura, K.; Nakamura, T. A concentric gravity wave structure in the mesospheric airglow images. J. Geophys. Res. Atmos. 2007, 112, D02102. [Google Scholar] [CrossRef]
  12. Lai, C.; Jia, Y.; Xu, J.; Yuan, W.; Li, Q.; Liu, X.; Lai, C.; Jia, Y.; Xu, J.; Yuan, W. Detection of large-scale concentric gravity waves from a Chinese airglow imager network. J. Atmos. Sol. -Terr. Phys. 2017, 30, 269–276. [Google Scholar] [CrossRef]
  13. Chi, W. Recent advances in observation and research of the Chinese Meridian Project. Chin. J. Space Sci. 2018, 38, 640–649. [Google Scholar]
  14. Lai, C.; Xu, J.; Yue, J.; Yuan, W.; Liu, X.; Li, W.; Li, Q. Automatic extraction of gravity waves from all-sky airglow image based on machine learning. Remote Sens. 2019, 11, 1516. [Google Scholar] [CrossRef]
  15. Miller, S.D.; Mills, S.P.; Elvidge, C.D.; Lindsey, D.T.; Lee, T.F.; Hawkins, J.D. Suomi satellite brings to light a unique frontier of nighttime environmental sensing capabilities. Proc. Natl. Acad. Sci. USA 2012, 109, 15706–15711. [Google Scholar] [CrossRef]
  16. Yue, J.; Hoffmann, L.; Alexander, M.J. Simultaneous observations of convective gravity waves from a ground-based airglow imager and the AIRS satellite experiment. J. Geophys. Res. Atmos. 2013, 118, 3178–3191. [Google Scholar] [CrossRef]
  17. Gupta, S.; Sameer, M.; Mohan, N. Detection of epileptic seizures using convolutional neural network. In Proceedings of the 2021 International Conference on Emerging Smart Computing and Informatics (ESCI), Pune, India, 5–7 March 2021; pp. 786–790. [Google Scholar]
  18. Hubel, D.H.; Wiesel, T.N. Receptive fields of single neurones in the cat’s striate cortex. J. Physiol. 1959, 148, 574. [Google Scholar] [CrossRef]
  19. LeCun, Y.; Bottou, L.; Bengio, Y.; Haffner, P. Gradient-based learning applied to document recognition. Proc. IEEE 1998, 86, 2278–2324. [Google Scholar] [CrossRef]
  20. Indolia, S.; Goswami, A.K.; Mishra, S.P.; Asopa, P. Conceptual understanding of convolutional neural network-a deep learning approach. Procedia Comput. Sci. 2018, 132, 679–688. [Google Scholar] [CrossRef]
  21. Chegeni, M.K.; Rashno, A.; Fadaei, S. Convolution-layer parameters optimization in convolutional neural networks. Knowl. Based Syst. 2023, 261, 110210. [Google Scholar] [CrossRef]
  22. Krizhevsky, A.; Sutskever, I.; Hinton, G.E. Imagenet classification with deep convolutional neural networks. Adv. Neural Inf. Process. Syst. 2012, 25, 84–90. [Google Scholar] [CrossRef]
  23. Szegedy, C.; Liu, W.; Jia, Y.; Sermanet, P.; Reed, S.; Anguelov, D.; Erhan, D.; Vanhoucke, V.; Rabinovich, A. Going deeper with convolutions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Boston, MA, USA, 7–12 June 2015; pp. 1–9. [Google Scholar]
  24. Wu, X.; Liu, R.; Yang, H.; Chen, Z. An xception based convolutional neural network for scene image classification with transfer learning. In Proceedings of the 2020 2nd International Conference on Information Technology and Computer Application (ITCA), Guangzhou, China, 18–20 December 2020; pp. 262–267. [Google Scholar]
  25. Simonyan, K.; Zisserman, A. Very deep convolutional networks for large-scale image recognition. arXiv 2014, arXiv:1409.1556. [Google Scholar]
  26. Bhosle, K.; Musande, V. Evaluation of deep learning CNN model for land use land cover classification and crop identification using hyperspectral remote sensing images. J. Indian Soc. Remote Sens. 2019, 47, 1949–1958. [Google Scholar] [CrossRef]
  27. Wang, C.; Mouche, A.; Tandeo, P.; Stopa, J.; Chapron, B.; Foster, R.; Vandemark, D. Automated geophysical classification of sentinel-1 wave mode sar images through deep-learning. In Proceedings of the IGARSS 2018-2018 IEEE International Geoscience and Remote Sensing Symposium, Valencia, Spain, 22–27 July 2018; pp. 1776–1779. [Google Scholar]
  28. Guo, C.; Ai, W.; Hu, S.; Du, X.; Chen, N. Sea surface wind direction retrieval based on convolution neural network and wavelet analysis. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2022, 15, 3868–3876. [Google Scholar] [CrossRef]
  29. Matsuoka, D.; Watanabe, S.; Sato, K.; Kawazoe, S.; Yu, W.; Easterbrook, S. Application of deep learning to estimate atmospheric gravity wave parameters in reanalysis data sets. Geophys. Res. Lett. 2020, 47, e2020GL089436. [Google Scholar] [CrossRef]
  30. Sreekanth, V.S.; Raghunath, K.; Mishra, D. Deep Kernel Dictionary Learning for detection of wave breaking features in Atmospheric Gravity Waves. Comput. Geosci. 2023, 176, 105361. [Google Scholar] [CrossRef]
  31. González, J.L.; Chapman, T.; Chen, K.; Nguyen, H.; Chambers, L.; Mostafa, S.A.; Wang, J.; Purushotham, S.; Wang, C.; Yue, J. Atmospheric Gravity Wave Detection Using Transfer Learning Techniques. In Proceedings of the 2022 IEEE/ACM International Conference on Big Data Computing, Applications and Technologies (BDCAT), Vancouver, WA, USA, 6–9 December 2022; pp. 128–137. [Google Scholar]
  32. Lee, S.; Chiang, K.; Xiong, X.; Sun, C.; Anderson, S. The S-NPP VIIRS Day-Night Band on-orbit calibration/characterization and current state of SDR products. Remote Sens. 2014, 6, 12427–12446. [Google Scholar] [CrossRef]
  33. Hillger, D.; Seaman, C.; Liang, C.; Miller, S.; Lindsey, D.; Kopp, T. Suomi NPP VIIRS imagery evaluation. J. Geophys. Res. Atmos. 2014, 119, 6440–6455. [Google Scholar] [CrossRef]
  34. Hillger, D.; Kopp, T.; Lee, T.; Lindsey, D.; Seaman, C.; Miller, S.; Solbrig, J.; Kidder, S.; Bachmeier, S.; Jasmin, T. First-Light Imagery from Suomi NPP VIIRS. Bull. Am. Meteorol. Soc. 2013, 94, 1019–1029. [Google Scholar] [CrossRef]
  35. Cao, C.; De Luccia, F.J.; Xiong, X.; Wolfe, R.; Weng, F. Early on-orbit performance of the visible infrared imaging radiometer suite onboard the suomi national polar-orbiting partnership (S-NPP) satellite. IEEE Trans. Geosci. Remote Sens. 2014, 52, 1142–1156. [Google Scholar] [CrossRef]
  36. Chang, S.G.; Yu, B.; Vetterli, M. Adaptive wavelet thresholding for image denoising and compression. IEEE Trans. Image Process. 2000, 9, 1532–1546. [Google Scholar] [CrossRef]
  37. Kyba, C.C.; Kuester, T.; Sánchez de Miguel, A.; Baugh, K.; Jechow, A.; Hölker, F.; Bennie, J.; Elvidge, C.D.; Gaston, K.J.; Guanter, L. Artificially lit surface of Earth at night increasing in radiance and extent. Sci. Adv. 2017, 3, e1701528. [Google Scholar] [CrossRef]
Electronics 13 04030 g001
Figure 1. Schematic diagram of transfer learning structure.
Figure 1. Schematic diagram of transfer learning structure.
Electronics 13 04030 g001
Electronics 13 04030 g002
Figure 2. Flowchart of the gravity wave detection algorithm for low-light images based on deep learning.
Figure 2. Flowchart of the gravity wave detection algorithm for low-light images based on deep learning.
Electronics 13 04030 g002
Electronics 13 04030 g003
Figure 3. Training process of the Inception V3 model.
Figure 3. Training process of the Inception V3 model.
Electronics 13 04030 g003
Electronics 13 04030 g004
Figure 4. Performance of YOLO v10 model.
Figure 4. Performance of YOLO v10 model.
Electronics 13 04030 g004
Table 1. Training results of the transfer learning model.
Table 1. Training results of the transfer learning model.
ModelAccuracy in %
TrainValidationTest
Inception V389.787.988.2
Table 2. Results of gravity wave classification detection.
Table 2. Results of gravity wave classification detection.
ModelPrecisionRecall
Inception V387.9%91.1%
YOLO v1046.7%50.1%
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Xiao, B.; Hu, S.; Ai, W.; Li, Y. Atmospheric Gravity Wave Detection in Low-Light Images: A Transfer Learning Approach. Electronics 2024, 13, 4030. https://doi.org/10.3390/electronics13204030

AMA Style

Xiao B, Hu S, Ai W, Li Y. Atmospheric Gravity Wave Detection in Low-Light Images: A Transfer Learning Approach. Electronics. 2024; 13(20):4030. https://doi.org/10.3390/electronics13204030

Chicago/Turabian Style

Xiao, Beimin, Shensen Hu, Weihua Ai, and Yi Li. 2024. "Atmospheric Gravity Wave Detection in Low-Light Images: A Transfer Learning Approach" Electronics 13, no. 20: 4030. https://doi.org/10.3390/electronics13204030

APA Style

Xiao, B., Hu, S., Ai, W., & Li, Y. (2024). Atmospheric Gravity Wave Detection in Low-Light Images: A Transfer Learning Approach. Electronics, 13(20), 4030. https://doi.org/10.3390/electronics13204030

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop