Ewha Womans Univ.
Application of Artificial Neural Network for the direct estimation of
atmospheric instability from a geostationary satellite imager
Su Jeong Lee, Myoung-Hwan Ahn, Yeonjin Lee
Department of Atmospheric Sciences and Engineering, Ewha Womans University
Introduction
Results and Discussion
Advanced Meteorological Imager (AMI)
Network training with IASI data
 AMI will be onboard next generation Geostationary Korea Multi-Purpose Satellite
(GEO-KOMPSAT-2A) in 2018
 Designed to have improved spectral, temporal, and spatial resolution compared to the
first generation imager, so many new and improved value-added products are
expected to be produced
 Among the new products, atmospheric instability information is one of the
important new possibilities with the pseudo-sounding capability of the imager, and
this information could provide a significant addition for the nowcasting activities:
Convective Available Potential Energy (CAPE) is selected for the study
Necessary vertical T and q profiles for
training are from the retrieved values
from IASI, obtained from EUMETSAT
To construct a uniform set of training
data, the same amount of data (2,399)
from three different CAPE ranges, total
7,197, was selected out of 63,877
profiles.
Approaches to retrieve instability information from Imagers
Physical retrieval
Statistical retrieval
Satellite observed radiances
Satellite observed radiances
infer (inversion technique)
directly estimate
using ANN
Alternative
Temperature/humidity profiles
Atmos. Stability indices (e.g.,CAPE)
Atmos. Stability indices (e.g.,CAPE)
Currently used for MSG and GOES
sounder to derive stability information
Using an Artificial Neural Network (ANN)
to directly derive stability information
from the measured radiances (or
brightness temperatures, Tbs)
• Computationally fast
• Can retain the original high spatial and
temporal resolution data
Limitation :
• Computer-intensive
• Need to reduce the (spatial/temporal)
resolution of the imagery data
(Koenig and Coning, 2008)
Data and Methodology
Objective
selection
CAPE
Final set of
CAPE
profiles
ANN training
MODTRAN run
CAPE Range
0 <= CAPE <1000
1000 <= CAPE <2000
2000 <= CAPE
TOTAL
63,877
(establish relations between
CAPE and Tb)
Simulated
Radiance (I)
TOTAL
selected
53,600
2,399
7,878
2,399
2,399
2,399
Combination of weights that results in the best network performance
Input layer
cir_day
cir_time
latitude
longitude
sat_zenith
TB6.2
TB7.3
TB8.7
TB9.6
TB10.8
TB12.0
TB13.4
INPUT
Total weights
( WIHxWHO )
circular day
circular time
Latitude
Longitude
0.09
0.22
0.02
0.32
Sat. zenith
-0.18
Tb6.2
Tb7.3
Tb8.7
Tb9.6
Tb10.8
Tb12.0
Tb13.4
-0.71
1.56
6.79
4.63
-0.38
-7.04
-4.17
Hidden layer
n1
n2
n3
Output
layer
n4
n5
n6
n7
n8
n9
n 10
n 11
n 12
CAPE
strongest weight
2nd strongest weight
* thickness of the arrows: relative magnitude of connection weights
Work Flow
T, q profiles
(IASI level2)
The combination of ANN parameters
that produces best algorithm
performance is:
- the number of hidden neuron: 12
- the number of epochs: 3000
- Learning rate: 0.3
- The analysis of final combination of weights reveals that among 12 input variables,
brightness temperatures, particularly Tb8.7 and Tb12.0, are identified as discriminating
input measurements, having strong connections with the neurons in the hidden layer
- Detailed connections for each variable need further investigation
CAPE derived using ANN algorithm
SEVIRI SRF
7,197
4,000
band-averaged
Radiance (𝑰)
determine
Inverse of Planck eq.
Theoretical TB
3,000
09:02:14
11:02:13
2,000
Best ANN Parameters
1,000
0
Measured
radiance (TB)
Validation
CAPE
direct estimation
SEVIRI 2.5min rapid scan
DATA
 IASI (Infrared Atmospheric Sounding Interferometer) Level 2 Measurements
13:02:12
15:02:15
- used for the simulation of SEVIRI Tb
- Temperature and humidity profiles from Jan. to Dec. in 2012 between 35N-75N latitude and
75W-75E longitude
 SEVIRI (Spinning Enhanced Visible and InfraRed Imager) 2.5min rapid scan data
- used for the characterization and validation of the ANN model
- Tbs of the seven IR channels, satellite zenith angle, lat/lon in June 20, 2013
17:02:14
17:02:14
Radiative Transfer Model
 Significant features
 MODTRAN (MODerate resolution TRANsmission) v. 5.2.2
- used for the simulation of the upwelling radiances at TOA
- Simulated radiances are band-averaged over SEVIRI spectral response function (SRF)
Artificial Neural Network (ANN)
 Multi-Layer Perceptron feedforward backpropagation Algorithm
- used to find the relation between the simulated Tb and CAPE and to retrieve CAPE directly
from the measured Tb
INPUT
layer
latitude
longitude
time
sat_zen
Tb 6.2
Tb 7.3
Tb 8.7
Tb 9.6
Tb 10.8
Tb 12.0
Tb 13.4
𝑥1
𝑥2
•
•
•
HIDDEN
layer
𝑤11
𝑤12
∑
1
𝑓(·)
𝑤13
∑
2
𝑓(·)
𝑥𝑛
𝑣1
𝑣2
•
•
•
𝑤n1
𝑤n2
𝑤nm
OUPUT
layer
∑
𝒙𝒊 : input , 𝑦: output
𝑦
𝒘𝒊𝒋 : connection strength(weight) between
input neuron 𝑖 hidden neuron 𝑗
𝒗𝒋 : connection strength between hidden
neuron 𝑗 and the output neuron
𝑓(·) : activation(transfer) function (hyperbolic
tangent)
𝒎
𝑣m
𝒚=
𝒏
𝒗𝒊 𝒇
𝒋=𝟏
∑ m 𝑓(·)
neuron
𝒘𝒊𝒋 𝒙𝒊
𝒊=𝟏
(Blackwell & Chen, 2009)
- network weights are updated in the direction in which the error (difference between the
desired output and the actual output) decreases most rapidly
- For every iteration (epoch), the network fitness is evaluated by root mean square error (RMSE)
𝑅𝑀𝑆𝐸 =
∑𝑝( 𝑦𝑡𝑎𝑟𝑔𝑒𝑡
𝑝
− 𝑦 (𝑝) )2
# 𝑜𝑓 𝑝𝑎𝑡𝑡𝑒𝑟𝑛
20:02:13
𝑝: 𝑝𝑎𝑡𝑡𝑒𝑟𝑛(𝑠𝑎𝑚𝑝𝑙𝑒 𝑠𝑖𝑧𝑒)
- During the afternoon hours (at around 13:02 UTC) several convective clouds begin to pop
up over the regions where the CAPE values are relatively high (marked with arrows and
circles at 09:02 UTC image) and developed to the severe convective clouds (at 15:02 UTC)
- High CAPE values around the leading edges of clouds induce a further development, while
weaker CAPE values around the trailing edges result in weakened convective activities.
Significance will be assessed with more case studies and quantitative validation.
- No significant convection occurs over high CAPE areas in the morning images (at around
11:02 UTC marked with dashed blue arrows) which requires further investigation
 Further study
- Apply the model to more special observation data and compare with CAPE values from
the radiosonde for the quantitative validation
References
Blackwell W. J. and F. W. Chen, 2009: Neural Networks in Atmospheric Remote Sensing, Massachusetts
Institute of Technology, MA, USA
Gardner, M. W. and Dorling, S. R., 1998: Artificial Neural Networks (the Multilayer Perceptron): a review
of applications in the atmospheric sciences. Atmos. Env., 32, pp2627-2636
Koenig, M. and E. D. Coning, 2008: The MSG Global Instability Indices Product and Its Use as a Now
casting Tool. Weather and Forecasting., 24, 272-285
Acknowledgement
We would like to give special thank to Marianne Koenig for all the technical support and guidance.
This research was funded by the National Meteorological Satellite Center (NMSC) titled “A Study on the
Development of Meteorological Data Processing Techniques of Geostationary Meteorological Satellite”