Korean Journal of Remote Sensing (대한원격탐사학회지)

Korean Society of Remote Sensing (대한원격탐사학회)
권호 표지
DOI QR코드

DOI QR Code

An Experiment on Image Restoration Applying the Cycle Generative Adversarial Network to Partial Occlusion Kompsat-3A Image

  • Won, Taeyeon (Department of Advanced Technology Fusion, Konkuk University) ;
  • Eo, Yang Dam (Department of Civil and Environmental Engineering, Konkuk University)
  • Received : 2022.01.26
  • Accepted : 2022.02.17
  • Published : 2022.02.28
Download PDF
⟨ Previous Next ⟩

Abstract

This study presents a method to restore an optical satellite image with distortion and occlusion due to fog, haze, and clouds to one that minimizes degradation factors by referring to the same type of peripheral image. Specifically, the time and cost of re-photographing were reduced by partially occluding a region. To maintain the original image's pixel value as much as possible and to maintain restored and unrestored area continuity, a simulation restoration technique modified with the Cycle Generative Adversarial Network (CycleGAN) method was developed. The accuracy of the simulated image was analyzed by comparing CycleGAN and histogram matching, as well as the pixel value distribution, with the original image. The results show that for Site 1 (out of three sites), the root mean square error and R2 of CycleGAN were 169.36 and 0.9917, respectively, showing lower errors than those for histogram matching (170.43 and 0.9896, respectively). Further, comparison of the mean and standard deviation values of images simulated by CycleGAN and histogram matching with the ground truth pixel values confirmed the CycleGAN methodology as being closer to the ground truth value. Even for the histogram distribution of the simulated images, CycleGAN was closer to the ground truth than histogram matching.

Keywords

1. Introduction

The number of days of optical imaging has drastically decreased due to the increase in cloud generation and air pollution caused by climate change, making it difficult to acquire images in a timely manner or to acquire periodic images necessary for time series analysis (Lee et al., 2017). Therefore, from the point of view of remote sensing image users, it is realistically highly important to increase the number of available images to improve the reliability of the results.

Radiometric (atmospheric) correction to normalize weather conditions at the time of acquiring images is an essential processing step for satellite images, and the types of radiometric correction can be divided into absolute and relative methods (Du et al., 2002). Absolute radiometric correction is impossible if there are no terrestrial observations at the time at which the satellite image is taken. However, in relative radiometric correction, if there is one image to be corrected and one image to be used as a reference, the correction proceeds with the data acquired by the same sensor or weather conditions through linear regression normalization or histogram matching.

Till date, simulation image production and radiometric and atmospheric correction techniques have been applied as alternatives to overcome the limitations of optical imaging conditions caused by bad weather; yet, only histogram matching for high resolution images is usually installed in commercial programs, so local distortion of the images cannot be avoided (Lee et al., 2017). Therefore, various relative radiation-correction techniques for high-resolution satellite images have been studied to solve these practical problems. Representations of these techniques include histogram matching, the PIF-based pixel value transformation method, and the method of performing relative radiometric correction through frequency domain transformation (Chen et al., 2018; Seo and Eo, 2018, 2019). However, if any part of the original image is completely occluded by dark shadows or clouds, it is necessary to consider a new method of importing or restoring the region from an alternative image. Various studies have been conducted to restore the occluded area. Yoo and Lee (2010) conducted a study to restore the occluded area using patch-based processing for aerial images and LiDAR (Light Detection and Ranging) data (Yoo and Lee, 2010). Su et al. (2016) conducted a study on shadow detection and shadow detection and removal of shadow-occluded high-resolution panchromatic images. After conducting cloud detection by combining bimodal histogram splitting and image matting techniques, they developed spatial adaptive nonlocal sparse technology. However, this technique was only able to restore uniform objects in the shadow area(Su et al., 2016). Theoretically, the post-processing process for restoration is inevitably complicated because the observation time and camera angle are different.

Therefore, in this study, a simulated image was generated by learning 4-band satellite images with a range of 9 to 16 bits instead of the 3-band and 8-bit images that were mainly used for deep learning in the CycleGAN system. By applying the proposed method to a Kompsat3A image, an experiment was conducted to restore the areas occluded by fog, haze, and clouds, and the accuracy of the simulation results was evaluated.

2. GAN (Generative Adversarial Network)

Within artificial intelligence, deep learning is active in many fields, and among them, the field related to images is sometimes called vision. In the existing vision field, image classification is performed through a fully connected network (FCN), which is a relatively inefficient neural network that has a high load even with hardware, because all nodes are connected. Because of this, a period of stagnation occurred in terms of research in the field of deep-learning vision, owing to high computational load and overfitting. However, this was alleviated when Lecun et al. (1989) proposed a neural network for classifying handwritten digits using a convolutional neural network (CNN), which uses feature extraction through convolutional means and downsampling through pooling. The use of the CNN ushered in a new era in the field of deep-learning vision because of the CNN’s higher accuracy, more reasonable computational load, and less tendency to overfit as compared to the existing FCN (Lecun et al., 1989).

Many studies have since been conducted to improve this CNN version, some of which have generated new outputs through the classification result rather than ending at the simple image classification stage. One method for this is the fully convolutional network (FCN), which takes the form of attaching the convolutional layer (the feature extraction part of the CNN) in the up and down direction; further, instead of the classifier used in the existing CNN, it uses a deconvolution layer that upsamples the image using the extracted features (Badrinarayanan et al., 2017). Many remote sensing studies for deep-learning neural networks have used FCN to learn not only input values as images, but also result values as images, in order to derive results in the form of images.

The GAN structure consists of two models: the generator (G) model, responsible for generation, and the discriminator (D) model, responsible for discriminating. These two models learn adversarially to each other; the G model generates an output that predicts whether the D model is true, and the D model learns to distinguish between the true and false results of the G model learned by real data. As a result, data with a sampling distribution similar to that of the true data can be generated using either a random vector z (called random noise), or latent variable, as an input. The two models G and D have their weights updated through a connected loss function, such that learning takes place interactively (Goodfellow et al., 2014). With the invention of the GAN algorithm, a new paradigm was presented, in which results are produced that imitate learning data in limited deep learning that classifies and detects an object.

After the development of GAN, a study related to inpainting using GAN was conducted, introducing the Deepfill algorithm (Yu et al., 2019). Deepfill is an algorithm that selects pixels for training images using gate convolution and then learns the features, making it possible to restore the features of the learned image related to the atypical mask and the surrounding features of the mask (Lee and Bae, 2021).

In addition, an image conversion method that produces an image by creating a relationship between specific datasets, rather than generating data based on simple input data, has been developed. Pix2Pix, a type of conditional GAN, aims to map the input and output images, and has the characteristic of pairing and training data between two datasets (Isola et al., 2017).

The CycleGAN method is an extension of the Pix2Pix methodology and was initially published in 2017. However, unlike Pix2Pix, which must be trained in pairs, CycleGAN has the advantage of being able to map unpaired learning. In this process, an input image is transformed into an output image through a model, and the concept of cyclic consistency is used for the language translation, which maps the output image to the original input image, as shown in Fig. 1. Because of the loss function using cyclic coherence, we were able to change the characteristics of each image well, while at the same time minimizing the loss and change between images (Zhu et al., 2017).

OGCSBN_2022_v38n1_33_f0001.png 이미지

Fig. 1. The network architecture of CycleGAN used in this experiment. 

Recently, with the development of artificial intelligence beyond the scope of classification and detection, various studies have been conducted to apply the Generative Adversarial Network (GAN) model, which generates output values with similar characteristics using existing data in the field ofremote sensing. This is one of the techniques supporting the argument that a learning-based approach is essentialfor restoring the full image area, rather than the limited restoration considering only the sub-area, because of the difficulty of geometric registration and shadows in the case of high-resolution satellite images(Guo et al., 2018; Zhang et al., 2018). In light of this, Choi et al. (2020) proposed a method to create an aerial image learning dataset using theCycle GenerativeAdversarial Network (CycleGAN), and Ghamisi and Yokoya conducted an experiment to convert an image to digital surface model data using a conditional GAN (Choi et al., 2020; Ghamisi and Yokoya, 2018). In addition, Andrade et al.(2020) conducted a study to simulate the historical map of the past with the current satellite image using the Pix2Pix methodology for image segmentation and conditional GAN (Andrade et al., 2020). Yuan et al. (2020) simulated a near-infrared image using conditional GAN and RGB (red-greenblue)images(Yuan et al., 2020). However, most ofthe satellite images used in deep learning are 8-bit images with pixel valuesranging from0 to 255 by compressing the original raster. This is relevant because most deeplearning models are trained by receiving three bands (R, G, and B) asinput. This poses a problem in that the near-infrared band possessed by most satellite images cannot be used.

3. Experiment and analysis

1) Experimental method

In the case of a mid-resolution image, there is no issue with position accuracy, but in the case of a high resolution image, accurate positioning is essential to correct and restore pixel values from a reference image. In particular, it is difficult to maintain geometrical consistency through relative registration between images because high-resolution images cannot satisfy the perfectly matching geometrical conditions for each image, even with the same type of sensor. The proposed method of this study aimed to overcome this problem by learning the relationship between two images without geometric registration. Initially, an experiment was also conducted to remove clouds using the Deepfill algorithm. Because the algorithm was newly generated based on the surrounding pixels and learned data that did not take into account the features under the clouds, visually good results could be obtained; however, they were ultimately dismissed because they were deemed unsuitable for the purpose of this study. In light of this, the CycleGAN methodology was selected instead of the Pix2Pix paired-learning approach, since it enables unpaired learning in which data are paired for high resolution satellite image simulation without geometric registration.

Because the existing CycleGAN algorithm can only receive general image formats (jpg, jpeg, and png), only eight bits with 0-255 values and three bands with only R, G, and B could be used as input data. However, the high-resolution satellite image selected for the experiment was a 14-bit image with values from 0 to 16383, and had four RGBN (red, green, blue, and near-infrared) bands. Therefore, to learn such a high resolution image, the data input, processing, and simulation image generation were modified. The components composed of the existing Python image library were removed, and the image load and save functions using Gdal of Osgeo were newly created and replaced with components. A component for handling satellite image metadata was also added, and the CycleGAN model changed the input and output parts from three input and output parts to four. In the case of the discriminator, a model based on the ResNet structure was used, and in the case of the generator, the ResnetGenerator class was applied.

In general, for deep-learning model training, if the learning rate is too high, the weight of the deep-learning model is rapidly updated, and learning is not performed normally, while the simulated image output produces a strange shape. If the learning rate is too small, the weight update occurs subtly, and the established relationship between the input image and the target image is insufficient. Therefore, in this experiment, hyperparameters were empirically set using a batch size of 2, learning rate of 0.0005, epoch of 150, and linear decay epoch of 100. In addition, the most important hyperparameter settings, such as pool size and epoch, were set empirically by the experiment itself.

2) Experimental data

For the experimental data, South Korean images captured by the KOMPSAT3Asatellite were used.The collection time, latitude, and longitude are summarized in Table 1. For two images of the same region at different times, the overlapping region was set as an ROI and extracted in a rectangularshape. Subsequently, the image was divided into patches of 512 × 512 pixels using the split raster tool of the ARCMAP program. By performing this processin three different regions, a total of three pairs of datasets were constructed: Site 1, Site 2, and Site 3 each have a pair of data A and data B. Approximately 80% of the images were used as training data, and approximately 20% were extracted and used as experimental data.

Table 1. The information of satellite image obtained using experiments

OGCSBN_2022_v38n1_33_t0001.png 이미지

As shown in Fig. 2, Site 1 consists of forest area and urban area, and dataset A is a clear image that does not contain clouds or aerosols. Dataset B, which contains clouds and aerosols, is shown in Fig. 3. Therefore, based on dataset B, the areas without clouds and aerosols were classified as the learning dataset, and the areas with clouds and aerosols visually were classified as the test dataset. For Site 2, since both periods A and B are images of a clear urban area, they were divided into learning and experimental datasets to facilitate accuracy evaluation. For Site 2, only urban areas existed, and there were no clouds or aerosols throughout the image. Site 3 is composed of rural areas, forests, and marine areas. And also, Site 3 didn’t have cloud or aerosols throughout the image.

OGCSBN_2022_v38n1_33_f0002.png 이미지

Fig. 2. Experimental region data: (a) data at 20190918 Fig. 3. Satellite image occluded by clouds and aerosols. (Site 1), (b) data at 20190917 (Site 1), (c) data at 20151211 (Site 2), (d) data at 20160108 (Site 2), (e) data at 20190308 (Site 3), (f) data at 20190412 (Site 3).

OGCSBN_2022_v38n1_33_f0003.png 이미지

Table 3. Comparison of the results for each site in the experiment

3) Comparison of the experimental results

Using the CycleGAN model, Train_A data and Train_B data with four bands ranging from 0 to 16383 were trained, and a Fake_B result was obtained by inputting Test_A data into the deep-learning model. Further, a comparison with the CycleGAN methodology was made using a widely-employed histogram matching technique. And in the case of histogram matching, the simulated data set A was obtained like the data set B by matching the data set A to the histogram of the data set B. For all experiments, ground truth was set as the test B dataset. In addition, because there is a limit to the specifications of the device to be studied, the image was divided into patches of 512 × 512 pixels. Because of this, the well-known accuracy evaluation indices PSNR (Peak Signal to noise ratio) and SSIM (Structural Similarity Index Measure) could not be used; therefore, the root means square error (RMSE) and R2were evaluated. First, because the CycleGAN algorithm used in this study is part of a deep-learning technique, there are various parameters that must be set for model learning. Accordingly, the pool size, which directly affects the image quality, and the accuracy improvement according to the epoch, were compared. As shown in Table 2, the highest R2score and lowest RMSE were obtained when the pool size was 12. When the pool size was lowered to approximately 8, the resulting simulated image looks as if it were smeared. In addition, when the pool size was increased to approximately 16, the simulated image produced a visually distinct sense of heterogeneity. Therefore, in the subsequent experiments, the pool size was fixed at 12 and the epoch, at 250, before the experiment was performed. In the case of Epoch, as mentioned above, a total of 250 repetitions were learned using 150 repetitions with a fixed learning rate and 100 repetitions with a linearly decreasing learning rate.

Table 2. Results of each pool size for Site 2

OGCSBN_2022_v38n1_33_t0002.png 이미지

The results of the learning and prediction for the three regions are shown in Table 3, and the simulated image is shown in Fig. 5. In Site 1, where the mountainous and urban areas are mixed, one of the two datasets has clouds and aerosols globally distributed; thus, in this case, the data used for learning were also used for prediction to evaluate the accuracy. Therefore, there is a possibility that the evaluation of Site 1 may not be quantitative compared to the case for the other regions. Nevertheless, as can be seen in Fig. 4, in the image simulation using CycleGAN, all four bands were able to obtain a histogram distribution closer to the ground truth than histogram matching. Table 4 shows the mean and standard deviation for each band for the Site 1 image, which shows that the image simulated using CycleGAN showed average and standard deviation values closer to the ground truth compared to histogram matching. In addition, it was confirmed that the RMSE between the image simulated by the cycleGAN method and the ground truth image was 169.36 and 0.9911 for R2, which was better than 170.43 and 0.9896 from the histogram matching method. It is considered that the image shows good results by selecting as the learning data mainly for areas where various features such as cities and mountains are included in the image and there are no aerosols and clouds with the visual. Accordingly, it was confirmed that the video simulation of the four bands performed well. However, owing to the 512 × 512-pixel patches, when mosaicing the simulated image, abnormalities occurred at the boundary between patches, as shown in Fig. 5.

Table 3. Comparison of the results for each site in the experiment

OGCSBN_2022_v38n1_33_t0003.png 이미지

Table 4. Distribution of data values according to the band for Site 1

OGCSBN_2022_v38n1_33_t0004.png 이미지

OGCSBN_2022_v38n1_33_f0004.png 이미지

Fig. 4. Histogram table of data values by band for Site 1.

OGCSBN_2022_v38n1_33_f0005.png 이미지

Fig. 5. Experimental results for Site 1: (a) simulated image from CycleGAN, (b) simulated image from histogram matching, (c) ground truth image, (d) pixel value abnormality at the boundary of the patch about simulated image from CycleGAN.

In the case of Site 2, which consists of urban areas, similar RMSE and higher R2values were obtained as compared to histogram matching with RMSE 171.83 and R20.9880; similarly, for Site 3, higher error and higher R2values were also obtained with RMSE 170.53, R20.9889. In this case, it is considered that the RMSE value is higher than that of histogram matching because the area of the sea has a negative effect on learning as the area includes the sea. These results can be explained as follows: unlike histogram matching, which adjusts the distribution of simple pixel values, deep learning establishes a relationship between the features of real A data and real B data and transforms it, so in this case, it is presumed that the transformed relationship showed a similar error and a higher correlation.

However, when the amount of data was insufficient, this had a negative effect, as shown in Fig. 6, in which grid-type pixel cracks can be observed on the simulated image. When the amount of data was sufficient, as in using 20 or more images, pixel cracking symptoms were not observed. Therefore, we conclude that if simulation of the occluded area is to be performed, it should be possible to accurately reflect the characteristics of sufficiently diverse cases by using 20 or more images. It was confirmed that learning must be performed again to simulate regions with different image characteristics and that lower errors can be obtained only if different hyperparameters are applied to each region.

OGCSBN_2022_v38n1_33_f0006.png 이미지

Fig. 6. Example of simulated image result obtained when the training dataset is insufficient.

4. Conclusion

Four-band 14-bit satellite image simulation performed using the CycleGAN model was tested, and the possibility of image restoration of areas occluded by fog, haze, and clouds was confirmed. The proposed method showed a lower error and higher correlation than the image correction result obtained by histogram matching. Unlike histogram matching, which performs correction through the distribution of simple pixel values, it can be judged that a better result is shown because the proposed method predicts a simulated image by generating a correlation between image features. The proposed method showed better R2score for all three experimental area images (Site 1, Site 2, and Site 3) with different regional characteristics; however, if the resulting image was mosaic and enlarged, it was divided into patches such that pixels were at the boundary between the patches, and in these cases, a discontinuity of values occurred, which should be overcome by increasing the image patch size. It is expected that better results will be obtained if additional training epoch is carried out in regions with various characteristics, or if the specifications of the experimental equipment of the device are sufficiently improved to enable learning and testing without image segmentation. In the future, the proposed method can be used to increase the accuracy of deep learning by conducting an experiment with very high-quality data, such as a true-ortho images, or by mixing the generated data with other satellite image data.

References

  1. Andrade, H.J. and B.J. Fernandes, 2020. Synthesis of Satellite-Like Urban Images From Historical Maps Using Conditional GAN, IEEE Geoscience and Remote Sensing Letters, 19:1-4. https://doi.org/10.1109/LGRS.2020.3023170
  2. Badrinarayanan, V., A. Kendall, and R. Cipolla, 2017. Segnet: A deep convolutional encoder-decoder architecture for image segmentation, IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(12): 2481-2495. https://doi.org/10.1109/TPAMI.2016.2644615
  3. Chen, Y., K. Sun, D. Li, T. Bai, and W. Li, 2018. Improved relative radiometric normalization method of remote sensing images for change detection, Journal of Applied Remote Sensing, 12(4): 045018.
  4. Choi, H.W., S.H. Lee, H.H. Kim, and Y.C. Suh, 2020. A Study on the Complementary Method of Aerial Image Learning Dataset Using Cycle Generative Adversarial Network, Journal of the Korean Society of Surveying, Geodesy, Photogrammetry and Cartography, 38(6): 499-509 (in Korean with English abstract). https://doi.org/10.7848/KSGPC.2020.38.6.499
  5. Du, Y., P.M. Teillet, and J. Chihlar, 2002. Radiometric normalization of multitemporal high-resolution satellite images with quality control for land cover change detection, Remote Sensing of Environment, 82(1): 123-134. https://doi.org/10.1016/S0034-4257(02)00029-9
  6. Ghamisi, P. and N. Yokoya, 2018. Img2dsm: Height simulation fromsingle imagery using conditional generative adversarial net, IEEE Geoscience and Remote Sensing Letters, 15(5): 794-798. https://doi.org/10.1109/lgrs.2018.2806945
  7. Goodfellow, I.,J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio, 2014. Generative adversarial nets, Advances in Neural Information Processing Systems, 27.
  8. Guo, Q., M. He, and A. Li, 2018. High-resolution remote-sensing image registration based on angle matching of edge point features, IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 11(8): 2881-2895. https://doi.org/10.1109/jstars.2018.2844295
  9. Helmer, E. H. and B. Ruefenacht, 2005. Cloud-free satellite image mosaics with regression trees and histogram matching, Photogrammetric Engineering & Remote Sensing, 71(9): 1079-1089. https://doi.org/10.14358/PERS.71.9.1079
  10. Isola, P., J.Y. Zhu, T. Zhou, and A. A. Efros, 2017. Image-to-image translation with conditional adversarial networks, Proc. of the 2017 IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, Hawaii,Jul. 22-25, pp. 1125-1134.
  11. Kim, H.J., D.K. Seo, Y.D. Eo, M.C. Jeon, and W.Y. Park, 2019.Multi-temporal nonlinear regression method for landsat image simulation, KSCE Journal of Civil Engineering, 23(2): 777-787. https://doi.org/10.1007/s12205-018-1157-5
  12. LeCun, Y., B. Boser, J.S. Denker, D. Henderson, R.E. Howard, W. Hubbard, and L.D. Jackel, 1989. Backpropagation applied to handwritten zip code recognition, Neural Computation, 1(4): 541-551. https://doi.org/10.1162/neco.1989.1.4.541
  13. Lee, J.M. and K.H. Bae, 2021.ApplicationTechnology of Image Inpainting Algorithm for Occlusion on Texture Image, Journal of Korean Society for GeospatialInformation Science, 29(4): 147-156 (in Korean with English abstract). https://doi.org/10.7319/kogsis.2021.29.4.147
  14. Lee, M.H., S.B. Lee, Y.D. Eo, S.W. Kim, J.H. Woo, and S.H. Han, 2017. A comparative study on generating simulated Landsat NDVI images using data fusion and regression method-the case of the Korean Peninsula, Environmental Monitoring and Assessment, 189(7): 1-13. https://doi.org/10.1007/s10661-016-5706-4
  15. Lee, S.B., W.Y. Park, Y.D. Eo, M.W. Pyeon, S. Han, S.H. Yeon, and B.K. Lee, 2017.Analysis on the applicability of simulated image from SPOT 4 HRVIR image, KSCE Journal of Civil Engineering, 21(4): 1434-1442. https://doi.org/10.1007/s12205-016-0522-5
  16. Seo, D.K. and Y.D. Eo, 2018. Relative radiometric normalization for high-spatial resolution satellite imagery based on multilayer perceptron, Journal of the Korean Society of Surveying, Geodesy, Photogrammetry and Cartography, 36(6): 515-523 (in Korean with English abstract). https://doi.org/10.7848/KSGPC.2018.36.6.515
  17. Seo, D.K. and Y.D. Eo, 2019. Local-Based Iterative Histogram Matching for Relative Radiometric Normalization, Journal of the Korean Society of Surveying, Geodesy, Photogrammetry and Cartography, 37(5): 323-330. https://doi.org/10.7848/ksgpc.2019.37.5.323
  18. Su, N., Y. Zhang, S. Tian, Y. Yan, and X. Miao, 2016. Shadow detection and removal for occluded object information recovery in urban high-resolution panchromatic satellite images, IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 9(6): 2568-2582. https://doi.org/10.1109/JSTARS.2016.2570234
  19. Yoo, E.J. and D.C. Lee, 2010. Patch-based processing and occlusion area recovery for true orthoimage generation, Journal of the Korean Society of Surveying, Geodesy, Photogrammetry and Cartography, 28(1): 83-92 (in Korean with English abstract).
  20. Yu,J., Z. Lin,J.Yang, X. Shen, X. Lu, andT. S. Huang, 2019. Free-form image inpainting with gated convolution, Proc. of the 2019 IEEE/CVF International Conference on Computer Vision, Long Beach, CA, Jun. 15-20, pp. 4471-4480.
  21. Yuan, X., J. Tian, and P. Reinartz, 2020. Generating artificial near infrared spectral band from rgb image using conditional generative adversarial network, ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 3: 279-285.
  22. Zhang, Q., Q. Yuan, C. Zeng, X. Li, and Y. Wei, 2018. Missing data reconstruction in remote sensing image with a unified spatial-temporal-spectral deep convolutional neural network, IEEE Transactions on Geoscience and Remote Sensing, 56(8): 4274-4288. https://doi.org/10.1109/tgrs.2018.2810208
  23. Zhou, H., S. Liu, J. He, Q. Wen, L. Song, and Y. Ma, 2016. A new model for the automatic relative radiometric normalization of multiple images with pseudo-invariant features, International Journal of Remote Sensing, 37(19): 4554-4573. https://doi.org/10.1080/01431161.2016.1213922
  24. Zhu, J.Y., T. Park, P. Isola, and A.A. Efros, 2017. Unpaired image-to-image translation using cycle-consistent adversarial networks, Proc. of the 2017 IEEE International Conference on Computer Vision, Venice, Italy, Oct. 22-29, pp. 2223-2232.