Retrieving Vertical Cloud Radar Reflectivity from MODIS Cloud Products with CGAN: An Evaluation for Different Cloud Types and Latitudes
<p>Implementation flowchart of the CGAN-based 3D cloud radar reflectivity retrieval technique. It is divided into a model training process and a 3D cloud map generation process. (MYD03 is a MODIS data file that provides detailed geolocation information, which is needed for comparison in the 3D cloud map generation step which is not discussed in this article. In this paper, we mainly discuss the model training process and the model evaluation.).</p> "> Figure 2
<p>The framework of the Leinonen et al. [<a href="#B20-remotesensing-15-00816" class="html-bibr">20</a>] CGAN model for retrieving vertical radar reflectivity based on the MODIS cloud products.</p> "> Figure 3
<p>Examples of the cloud vertical radar reflectivity generated by CGAN for the eight main cloud types. Each column presents a typical case of one cloud type. The top two rows of subplots in each column are the cloud observations from MODIS (matching the CloudSat cloud radar observation tracks) including CWP, <span class="html-italic">t<sub>c</sub></span>, <span class="html-italic">r</span><sub>e</sub>, and <span class="html-italic">P</span><sub>top</sub>. The third row is the vertical radar reflectivity generated based on CGAN and MODIS data (noise input is 0), the fourth row corresponds to the actual CloudSat radar reflectivity, and the last row is the corresponding cloud classification labels. Some of the cloud body labels cover a slightly larger area than the displayed reflectivity image because the reflectivity image does not include reflectivity between [−35 dBZ, −27 dBZ] to avoid noise interference.</p> "> Figure 4
<p>Multi-threshold multi-metric box plot of the test set. The horizontal axis is the four-evaluation metrics and the vertical axis is the score of each metric. The scores were calculated with five thresholds, shown in five colors. The outlier points are marked with diamond dots.</p> "> Figure 5
<p>POD vs. size scatterplot of all-sample in the test set for thresholds of −25 dBZ (<b>a1</b>–<b>h1</b>) and 0 (<b>a2</b>–<b>h2</b>). The horizontal coordinate is the cloud size (i.e., the number of pixels covered by each inverse cloud sample region) and the vertical coordinate is the POD score. A point in each subplot indicates the POD result of a retrieved cloud sample. The color shades represent the sample distribution density heat maps.</p> "> Figure 6
<p>The distribution of POD scores for all test samples with threshold −25 dBZ (<b>a1</b>–<b>h1</b>) and 0 dBZ (<b>a2</b>–<b>h2</b>) for the 8 cloud types. POD values of 0 and 1 are plotted separately to indicate the two special cases of complete missing and complete forecasting, respectively. The percentage of POD falling in each interval is the vertical coordinate. “Number of samples” represents the total number of samples used to plot the statistics.</p> "> Figure 7
<p>Same as <a href="#remotesensing-15-00816-f005" class="html-fig">Figure 5</a>, but for the −5 dBZ threshold and the results of the lower, mid, and high latitude are distinguished by different colors. The curves corresponding to the colors are the least-square fitted lines.</p> "> Figure 8
<p>Normalized reflectivity–altitude distribution. The top (<b>a</b>–<b>c</b>), middle (<b>d</b>–<b>f</b>), and bottom (<b>g</b>–<b>i</b>) panels are for lower, mid, and high latitudes, respectively. The left (<b>a</b>,<b>d</b>,<b>g</b>), middle (<b>b</b>,<b>e</b>,<b>h</b>), and right (<b>c</b>,<b>f</b>,<b>i</b>) panels are the result of the real dataset, the generated dataset, and the difference between the two, respectively.</p> ">
Abstract
:1. Introduction
2. Data
2.1. CloudSat Data
2.2. MODIS Data
3. Experimental Design and Model Introduction
3.1. Experimental Design
3.2. Data Pre-Processing
- (1)
- Matching the MODIS data: MOD06-AUX data are processed the same way as CloudSat data horizontally, both are intercepted into units of length 64 for this study. The 2B_GEOPROF data are divided into 64 slices vertically with a resolution of about 240 m. We selected 2B_GEOPROF with an altitude of about 700 m to 15,820 m for the model training. Each unit is referred to as a “scene” throughout this paper (following L19 [20]).
- (2)
- Information filtering: Use the information within the cloudmask of MODIS data to filter by day/night: day, surface type: sea, cloud mask status: determined.
- (3)
- Mask screening: The cloudmask is set to 1 when there is a cloud (Confident Cloudy by MODIS) and 0 for the rest of the cases, and the cloudmask of each pixel is also set to 0 when there is any MODIS variable with missing measurements. With this premise, if there are more than 32 cloudmask values of 0 in the 64 pixels of data, the data will be screened out. In order not to interfere with the training, we also screen out the cases where cloudmask marks clouds but there are almost no clouds.
- (4)
- After steps (2) and (3), 162,125 samples were deleted and only 64,616 samples were retained. Then, we normalized the reflectivity data of 2B_GEOPROF data from [−27 dBZ, 20 dBZ] to [−1, 1], and the reflectivity smaller than −27 dBZ was set to −1 (because the reflectivity data of 2B_GEOPROF are disturbed by clutter and exhibit mosaic traits, which need to be filtered out). L19 suggested a normalization range of [−35 dBZ, 20 dBZ] in their public dataset, the normalized data are free of noise. We find that if normalization is performed with −35 dBZ as the minimum value, there are many noises, namely mosaic-like weak echoes in the cloud-free regions. Therefore, we adjusted the normalization using [−27 dBZ, 20 dBZ], which is effective to avoid noise interference. When normalizing the MODIS data, the missing data are substituted with a value of 0. The normalization for different variables is calculated by entering the following equation (Equations (2)−(5) are replicated from L19 for referring convenience):
- (5)
- After the dataset construction, we divided the global datasets by latitude into higher latitude (latitude > 65), mid-latitude (latitude between 20 and 65), and lower latitude regions (latitude < 20) for a comparison study of the model retrievals over different latitude regions.
3.3. CGAN Convolutional Neural Network Model
3.4. Training Set and Test Set
3.5. Evaluation Index of Retrieval Results
- (1)
- Threat score (TS) measures the fraction of observed and/or forecast events that were correctly predicted. It can be thought of as the accuracy when correct negatives have been removed from consideration, that is, TS is only concerned with forecasts that count. It is calculated as follows:
- (2)
- False alarm rate (FAR) indicates the proportion of actual cloud-free areas in the prediction to the total predicted cloud area, which is calculated as follows:
- (3)
- The probability of detection (POD) indicates the proportion of the predicted cloud occupying the TRUE cloud area, which is calculated as follows:
- (4)
- HSS (Heidke’s skill score) indicates the accuracy of model forecasts after removing chance events, which can reflect the accuracy relative to the skill of random forecasts in a comprehensive manner. HSS is related to the threshold value, and a larger sample size is generally recommended. A larger HSS indicates better forecasts. A value of 0 indicates no skill, and HSS = 1 when the forecast is completely correct. Its calculation formula is
4. Retrieval Results Testing and Evaluation
4.1. Case Analysis
4.2. Statistical Verification Results
4.3. Statistical Evaluation for Different Cloud Types
4.4. Comparison of Cloud Retrievals at Different Latitudes
5. Conclusions and Discussion
- (1)
- The CGAN model can retrieve relatively realistic cloud vertical radar reflectivity structures, as well as cloud height and overall morphology, from MODIS satellite observations. The analysis of box plots with five cloud radar reflectivity thresholds (−25, −15, −5, 5, and 15 dBZ) and four metrics (TS, FAR, POD, and HSS) shows that the model possesses good skills for the low-reflectivity threshold representing the entire cloud bodies, but poor ones for recovering the high-reflectivity cloud cores. This method can identify more than 50% of the cloud areas that are close to the observations, but there exist about 30% false alarms on average. The model can retrieve about 25% of the medium-intensity cloud cores of the observation accurately. For the strongest reflectivity cores, the model is only slightly better than the completely random prediction. There are three possible explanations for this result. Firstly, the strong-echo cores are small, with a low frequency of occurrence, i.e., too few samples to train the model properly. Secondly, CloudSat observations are affected by signal attenuation as well as the complexity of the intense clouds, and thirdly, the model is trained without distinguishing strong and weak echoes, making the contribution of the strong echo unintendedly less weighted. We will try to improve the representation of the loss function or apply a self-attention block to improve it in the future.
- (2)
- The CGAN model performance was evaluated for eight cloud types individually. The CGAN model performs the best for deep convective clouds, with more than 60% of the sample cases having POD scores greater than 0.8 at a threshold of −25 dBZ, and more than 50% of the cases having POD scores greater than 0.6 at a threshold of 0 dBZ. The model also exhibits a good capability for retrieving nimbostratus (Ns). About 50% of the samples have POD scores greater than 0.8 at a threshold of −25 dBZ, and more than 40% of the samples have POD scores greater than 0.4 at a threshold of 0 dBZ, just slightly worse than those for the deep convective clouds. The model also demonstrates some retrieval ability for altostratus and cumulus, with about 60% of the test samples having PODs greater than 0.5 at a threshold of −25 dBZ, and more than 60% of the samples having PODs greater than 0.5 at a threshold of 0 dBZ. The method is less effective for Sc, Cirrus, altocumulus, and St, mainly because MODIS and CloudSat have weaker or less consistent observations of these clouds. In general, the model is effective for clouds with regular structures, large thicknesses, and strong reflectivity (large particles).
- (3)
- The cloud systems at the lower latitudes are taller and denser than those at the middle and high latitudes, and the CGAN model works the best. The model performance decreases toward high latitudes, roughly worsening by 10% in the mid latitude, and a further 25% in the high latitude. The CGAN model performed the best for deep convective clouds for all latitudes and with good skill scores. For Ns, a similar effect is achieved by the CGAN model for the mid latitude and low latitude regions. The retrievals of cumulus and altocumulus clouds in the mid latitude perform slightly better than the low latitude regions. The CGAN model exhibits an overall good ability to recover the CloudSat observed occurrence frequency distribution according to reflectivity magnitudes and heights and their variations with the latitude. The model overestimates the reflectivity intensity in the upper regions and underestimates the intensity of the strong core at the lower layer, and these errors are the smallest at low latitudes and increase with latitude.
Author Contributions
Funding
Data Availability Statement
Conflicts of Interest
References
- Werdell, P.; Behrenfeld, M.; Bontempi, P.; Boss, E.; Cairns, B.; Davis, G.; Franz, B.; Gliese, U.; Gorman, E.; Hasekamp, O.; et al. The Plankton, Aerosol, Cloud, Ocean Ecosystem Mission: Status, Science, Advances. Bull. Am. Meteorol. Soc. 2019, 100, 1775–1794. [Google Scholar] [CrossRef]
- Bocquet, M.; Elbern, H.; Eskes, H.; Hirtl, M.; Žabkar, R.; Carmichael, G.R.; Flemming, J.; Inness, A.; Pagowski, M.; Pérez Camaño, J.L.; et al. Data Assimilation in Atmospheric Chemistry Models: Current Status and Future Prospects for Coupled Chemistry Meteorology Models. Atmos. Chem. Phys. 2015, 15, 5325–5358. [Google Scholar] [CrossRef] [Green Version]
- Platnick, S.; Meyer, K.G.; King, M.D.; Wind, G.; Amarasinghe, N.; Marchant, B.; Arnold, G.T.; Zhang, Z.; Hubanks, P.A.; Holz, R.E.; et al. The MODIS Cloud Optical and Microphysical Products: Collection 6 Updates and Examples From Terra and Aqua. IEEE Trans. Geosci. Remote Sens. 2017, 55, 502–525. [Google Scholar] [CrossRef] [PubMed] [Green Version]
- Zhou, T.; Huang, Z.; Huang, J.; Li, J.; Bi, J.; Zhang, W. Study of Vertical Distribution of Cloud over Loess Plateau Based on a Ground based Lidar System. J. Arid. Meteorol. 2013, 2, 246–253+271. (In Chinese) [Google Scholar]
- Marshak, A.; Martins, J.V.; Zubko, V.; Kaufman, Y.J. What Does Reflection from Cloud Sides Tell Us about Vertical Distribution of Cloud Droplet Sizes? Atmos. Chem. Phys. 2006, 6, 5295–5305. [Google Scholar] [CrossRef] [Green Version]
- Hilburn, K.A.; Ebert-Uphoff, I.; Miller, S.D. Development and Interpretation of a Neural-Network-Based Synthetic Radar Reflectivity Estimator Using GOES-R Satellite Observations. J. Appl. Meteorol. Climatol. 2021, 60, 3–21. [Google Scholar] [CrossRef]
- Barker, H.W.; Jerg, M.P.; Wehr, T.; Kato, S.; Donovan, D.P.; Hogan, R.J. A 3D Cloud-Construction Algorithm for the EarthCARE Satellite Mission. Q. J. R. Meteorol. Soc. 2011, 137, 1042–1058. [Google Scholar] [CrossRef] [Green Version]
- Ham, S.; Kato, S.; Barker, H.W.; Rose, F.G.; Sun-Mack, S. Improving the Modelling of Short-wave Radiation through the Use of a 3D Scene Construction Algorithm. Q. J. R. Meteorol. Soc. 2015, 141, 1870–1883. [Google Scholar] [CrossRef]
- Zinner, T.; Marshak, A.; Lang, S.; Martins, J.V.; Mayer, B. Remote Sensing of Cloud Sides of Deep Convection: Towards a Three-Dimensional Retrieval of Cloud Particle Size Profiles. Atmos. Chem. Phys. 2008, 8, 4741–4757. [Google Scholar] [CrossRef] [Green Version]
- Marchand, R.; Mace, G.G.; Ackerman, T.; Stephens, G. Hydrometeor Detection Using Cloudsat—An Earth-Orbiting 94-GHz Cloud Radar. J. Atmos. Ocean. Technol. 2008, 25, 519–533. [Google Scholar] [CrossRef]
- Stephens, G.L.; Vane, D.G.; Tanelli, S.; Im, E.; Durden, S.; Rokey, M.; Reinke, D.; Partain, P.; Mace, G.G.; Austin, R.; et al. CloudSat Mission: Performance and Early Science after the First Year of Operation. J. Geophys. Res. 2008, 113, D00A18. [Google Scholar] [CrossRef]
- Minnis, P.; Sun-Mack, S.; Young, D.F.; Heck, P.W.; Garber, D.P.; Chen, Y.; Spangenberg, D.A.; Arduini, R.F.; Trepte, Q.Z.; Smith, W.L.; et al. CERES Edition-2 Cloud Property Retrievals Using TRMM VIRS and Terra and Aqua MODIS Data—Part I: Algorithms. IEEE Trans. Geosci. Remote Sens. 2011, 49, 4374–4400. [Google Scholar] [CrossRef]
- Zhang, C.; Chen, X.; Shao, H.; Chen, S.; Liu, T.; Chen, C.; Ding, Q.; Du, H. Evaluation and Intercomparison of High-Resolution Satellite Precipitation Estimates—GPM, TRMM, and CMORPH in the Tianshan Mountain Area. Remote Sens. 2018, 10, 1543. [Google Scholar] [CrossRef] [Green Version]
- Wang, S.; Han, Z.; Yao, Z. Comparison of Cloud Amounts from lSCCP and CloudSat over China and Its Neighborhood. Chin. J. Atmos. Sci. 2010, 34, 767–779. (In Chinese) [Google Scholar]
- Matrosov, S.Y. CloudSat Measurements of Landfalling Hurricanes Gustav and Ike (2008). J. Geophys. Res. 2011, 116, D01203. [Google Scholar] [CrossRef] [Green Version]
- Honda, T.; Miyoshi, T.; Lien, G.-Y.; Nishizawa, S.; Yoshida, R.; Adachi, S.A.; Terasaki, K.; Okamoto, K.; Tomita, H.; Bessho, K. Assimilating All-Sky Himawari-8 Satellite Infrared Radiances: A Case of Typhoon Soudelor (2015). Mon. Weather Rev. 2018, 146, 213–229. [Google Scholar] [CrossRef]
- Hua, J.; Wang, Z.; Duan, J.; Li, L.; Zhang, C.; Wu, X.; Fan, Q.; Chen, R.; Sun, X.; Zhao, L.; et al. Review of Geostationary Interferometric Infrared Sounder. Chin. Opt. Lett. 2018, 16, 111203. [Google Scholar] [CrossRef] [Green Version]
- Kotarba, A.Z. Calibration of Global MODIS Cloud Amount Using CALIOP Cloud Profiles. Atmos. Meas. Tech. 2020, 13, 4995–5012. [Google Scholar] [CrossRef]
- Goodfellow, I.; Pouget-Abadie, J.; Mirza, M.; Xu, B.; Warde-Farley, D.; Ozair, S.; Courville, A.; Bengio, Y. Generative Adversarial Networks. Commun. ACM 2020, 63, 139–144. [Google Scholar] [CrossRef]
- Leinonen, J.; Guillaume, A.; Yuan, T. Reconstruction of Cloud Vertical Structure with a Generative Adversarial Network. Geophys. Res. Lett. 2019, 46, 7035–7044. [Google Scholar] [CrossRef] [Green Version]
- Sassen, K.; Wang, Z. The Clouds of the Middle Troposphere: Composition, Radiative Impact, and Global Distribution. Surv. Geophys. 2012, 33, 677–691. [Google Scholar] [CrossRef]
- Sassen, K.; Wang, Z. Classifying Clouds around the Globe with the CloudSat Radar: 1-Year of Results. Geophys. Res. Lett. 2008, 35, L04805. [Google Scholar] [CrossRef]
- Mirza, M.; Osindero, S. Conditional Generative Adversarial Nets. arXiv 2014, arXiv:1411.1784. [Google Scholar]
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. |
© 2023 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/licenses/by/4.0/).
Share and Cite
Wang, F.; Liu, Y.; Zhou, Y.; Sun, R.; Duan, J.; Li, Y.; Ding, Q.; Wang, H. Retrieving Vertical Cloud Radar Reflectivity from MODIS Cloud Products with CGAN: An Evaluation for Different Cloud Types and Latitudes. Remote Sens. 2023, 15, 816. https://doi.org/10.3390/rs15030816
Wang F, Liu Y, Zhou Y, Sun R, Duan J, Li Y, Ding Q, Wang H. Retrieving Vertical Cloud Radar Reflectivity from MODIS Cloud Products with CGAN: An Evaluation for Different Cloud Types and Latitudes. Remote Sensing. 2023; 15(3):816. https://doi.org/10.3390/rs15030816
Chicago/Turabian StyleWang, Fengxian, Yubao Liu, Yongbo Zhou, Rongfu Sun, Jing Duan, Yang Li, Qiuji Ding, and Haoliang Wang. 2023. "Retrieving Vertical Cloud Radar Reflectivity from MODIS Cloud Products with CGAN: An Evaluation for Different Cloud Types and Latitudes" Remote Sensing 15, no. 3: 816. https://doi.org/10.3390/rs15030816