Machine Learning for Retrieving Cloud Optical Thickness ...
Transcript of Machine Learning for Retrieving Cloud Optical Thickness ...
![Page 1: Machine Learning for Retrieving Cloud Optical Thickness ...](https://reader033.fdocuments.in/reader033/viewer/2022060520/629657f5b5b5e71a2216046d/html5/thumbnails/1.jpg)
Machine Learning for Retrieving Cloud Optical Thickness from
Observed Reflectance: 3D EffectsCyberTraining 2020 Team 5
Kallista Angeloff1, Kirana Bergstrom2, Tianhao Le3, Chengtao Xu4
RA: Chamara Rajapakshe5, Kevin Zheng5, Mentor: Zhibo Zhang5
1 Department of Atmospheric Sciences, University of Washington2 Department of Mathematical and Statistical Sciences, University of Colorado Denver
3 Division of Geological and Planetary Sciences, California Institute of Technology 4 Department of Electrical Engineering, Embry-Riddle Aeronautical University, Daytona Beach
5Department of Physics, University of Maryland, Baltimore County
http://cybertraining.umbc.edu/
![Page 2: Machine Learning for Retrieving Cloud Optical Thickness ...](https://reader033.fdocuments.in/reader033/viewer/2022060520/629657f5b5b5e71a2216046d/html5/thumbnails/2.jpg)
Cloud optical thickness (COT)
• Vertical optical depth
• Determines how much solar radiation reaches the planet’s surface vs. how much is reflected or scattered
• Difficult to measure directly – radar, lidar, pyranometers
• Calculated (“retrieved”) from satellite observations of reflectance
CyberTraining: Big Data + High-Performance Computing + Atmospheric Sciences
![Page 3: Machine Learning for Retrieving Cloud Optical Thickness ...](https://reader033.fdocuments.in/reader033/viewer/2022060520/629657f5b5b5e71a2216046d/html5/thumbnails/3.jpg)
Retrieval algorithm and assumptions
• Underlying assumptions: plane-parallel (2D) clouds, each pixel independent
• Actual clouds: 3D effects – horizontal transport of radiation
CyberTraining: Big Data + High-Performance Computing + Atmospheric Sciences
![Page 4: Machine Learning for Retrieving Cloud Optical Thickness ...](https://reader033.fdocuments.in/reader033/viewer/2022060520/629657f5b5b5e71a2216046d/html5/thumbnails/4.jpg)
Synthetic data• 1D fractal cloud profiles – “lines of cloud”
CyberTraining: Big Data + High-Performance Computing + Atmospheric Sciences
![Page 5: Machine Learning for Retrieving Cloud Optical Thickness ...](https://reader033.fdocuments.in/reader033/viewer/2022060520/629657f5b5b5e71a2216046d/html5/thumbnails/5.jpg)
Synthetic data• Three connected parameters:
• Liquid water path (LWP), a measure of the weight of liquid water between two points in the atmosphere – fractal-like random variation
• Cloud drop effective radius (CER), a measure of the size distribution of water drops within a cloud – fixed; randomly assigned to each profile
• Cloud optical thickness (COT) – calculated from other two
CyberTraining: Big Data + High-Performance Computing + Atmospheric Sciences
![Page 6: Machine Learning for Retrieving Cloud Optical Thickness ...](https://reader033.fdocuments.in/reader033/viewer/2022060520/629657f5b5b5e71a2216046d/html5/thumbnails/6.jpg)
3D Radiative Transfer Model (SHDOM)
• Spherical Harmonic Discrete Ordinate Method (SHDOM)
• Inputs:• COT, CER, LWP• Solar zenith angle (SZA) – angle of the sunlight striking the cloud; 60°
• Outputs:• Reflectance
CyberTraining: Big Data + High-Performance Computing + Atmospheric Sciences
![Page 7: Machine Learning for Retrieving Cloud Optical Thickness ...](https://reader033.fdocuments.in/reader033/viewer/2022060520/629657f5b5b5e71a2216046d/html5/thumbnails/7.jpg)
3D Radiative Transfer Model (SHDOM)
• “Step cloud” check
CyberTraining: Big Data + High-Performance Computing + Atmospheric Sciences
![Page 8: Machine Learning for Retrieving Cloud Optical Thickness ...](https://reader033.fdocuments.in/reader033/viewer/2022060520/629657f5b5b5e71a2216046d/html5/thumbnails/8.jpg)
How to get from reflectance to COT?
• Historically an inverse problem
• Bispectral method – still used by MODIS, for example• One visible wavelength, one absorbing wavelength
• Simplification of machine learning: pattern recognition (statistical inference) rather than inversion
CyberTraining: Big Data + High-Performance Computing + Atmospheric Sciences
![Page 9: Machine Learning for Retrieving Cloud Optical Thickness ...](https://reader033.fdocuments.in/reader033/viewer/2022060520/629657f5b5b5e71a2216046d/html5/thumbnails/9.jpg)
Deep neural network (DNN)
reflectancechannels
DNN
COT
dx dx dx
CyberTraining: Big Data + High-Performance Computing + Atmospheric Sciences
![Page 10: Machine Learning for Retrieving Cloud Optical Thickness ...](https://reader033.fdocuments.in/reader033/viewer/2022060520/629657f5b5b5e71a2216046d/html5/thumbnails/10.jpg)
DNN structure
• Current DNN structure based on
DNN-2r
[Okamura et al 2017]
• Optimizer: Adam, Loss =
mean_square_error
• Batch = 4, epochs =10
Reflectance 4096 x 3
Fully Connected 8, relu
Fully Connected 8, relu
Fully Connected 8, relu
Fully Connected 8, relu
Fully Connected 1, Linear
COT 4096 x 1
CyberTraining: Big Data + High-Performance Computing + Atmospheric Sciences
![Page 11: Machine Learning for Retrieving Cloud Optical Thickness ...](https://reader033.fdocuments.in/reader033/viewer/2022060520/629657f5b5b5e71a2216046d/html5/thumbnails/11.jpg)
Results from deep neural network (DNN)
COT
CyberTraining: Big Data + High-Performance Computing + Atmospheric Sciences
![Page 12: Machine Learning for Retrieving Cloud Optical Thickness ...](https://reader033.fdocuments.in/reader033/viewer/2022060520/629657f5b5b5e71a2216046d/html5/thumbnails/12.jpg)
Results from deep neural network (DNN)
• Although MSE converged to a relatively low value, overall COT predictions were inaccurate (unphysical)
• Accuracy decreased with increasing spatial position
• Most likely explanation: overfitting
CyberTraining: Big Data + High-Performance Computing + Atmospheric Sciences
![Page 13: Machine Learning for Retrieving Cloud Optical Thickness ...](https://reader033.fdocuments.in/reader033/viewer/2022060520/629657f5b5b5e71a2216046d/html5/thumbnails/13.jpg)
Convolutional neural network (CNN) spatial slicing
halo halopoints of interest
reflectancechannels
CNN
COT
dx dx dx
CyberTraining: Big Data + High-Performance Computing + Atmospheric Sciences
![Page 14: Machine Learning for Retrieving Cloud Optical Thickness ...](https://reader033.fdocuments.in/reader033/viewer/2022060520/629657f5b5b5e71a2216046d/html5/thumbnails/14.jpg)
Convolutional neural network (CNN) spatial slicing
halo halopoints of interest
reflectancechannels
CNN
COT
dx dx dx
CyberTraining: Big Data + High-Performance Computing + Atmospheric Sciences
![Page 15: Machine Learning for Retrieving Cloud Optical Thickness ...](https://reader033.fdocuments.in/reader033/viewer/2022060520/629657f5b5b5e71a2216046d/html5/thumbnails/15.jpg)
Convolutional neural network (CNN) spatial slicing
halo halopoints of interest
reflectancechannels
CNN
COT
dx dx dx
CyberTraining: Big Data + High-Performance Computing + Atmospheric Sciences
![Page 16: Machine Learning for Retrieving Cloud Optical Thickness ...](https://reader033.fdocuments.in/reader033/viewer/2022060520/629657f5b5b5e71a2216046d/html5/thumbnails/16.jpg)
CNN structure
• Current CNN structure based on
DNN-4w [Okamura et al]
• Parallelization possible over spatial
slices
Reflectances 12 x 3
Convolutional 6, 100
Convolutional 1, 4
Dropout
Fully Connected 8
CyberTraining: Big Data + High-Performance Computing + Atmospheric Sciences
![Page 17: Machine Learning for Retrieving Cloud Optical Thickness ...](https://reader033.fdocuments.in/reader033/viewer/2022060520/629657f5b5b5e71a2216046d/html5/thumbnails/17.jpg)
Results from CNN
• Loss: mean squared logarithmic error
• Dropout rate: 0.5• Batch size: 1024• Epochs: 30
CyberTraining: Big Data + High-Performance Computing + Atmospheric Sciences
![Page 18: Machine Learning for Retrieving Cloud Optical Thickness ...](https://reader033.fdocuments.in/reader033/viewer/2022060520/629657f5b5b5e71a2216046d/html5/thumbnails/18.jpg)
Results from CNN
• Loss: mean squared error• Dropout rate: 0.5• Batch size: 1024• Epochs: 30
CyberTraining: Big Data + High-Performance Computing + Atmospheric Sciences
![Page 19: Machine Learning for Retrieving Cloud Optical Thickness ...](https://reader033.fdocuments.in/reader033/viewer/2022060520/629657f5b5b5e71a2216046d/html5/thumbnails/19.jpg)
Discussion of results
• The CNN was more successful than the DNN: why?• Not a one-to-one comparison – full data vs. spatial slicing• Nature of the data more closely resembles an image problem
• Edge (boundary) cases inherently more difficult
• More successful at predicting local rather than global COT• Some outliers and unphysical predictions – pre- or post-processing
CyberTraining: Big Data + High-Performance Computing + Atmospheric Sciences
![Page 20: Machine Learning for Retrieving Cloud Optical Thickness ...](https://reader033.fdocuments.in/reader033/viewer/2022060520/629657f5b5b5e71a2216046d/html5/thumbnails/20.jpg)
Future directions
• Expanded dataset – tuning • More data to better cover parameter range• More parameter tuning and structural changes
• Emulating a multi-scale model • Using a DNN-based global method to inform the local CNN
• 2D data – testing on satellite data, e.g. MODIS• Structure of spatially sliced CNN lends itself to parallelization
CyberTraining: Big Data + High-Performance Computing + Atmospheric Sciences