Upload
madison-gallagher
View
220
Download
0
Tags:
Embed Size (px)
Citation preview
1
Towards Automated Detection of Gulf Stream
North Wall From Concurrent Satellite
Images
Avijit GangopadhyayJeffrey Rezendes
Kevin LydonRamprasad Balasubramanian
Iren Valova
Outline
• Feature Oriented Regional Modeling System (FORMS) for the Western North Atlantic
• Gulf Stream path – issues• Manual Extraction process• Using SSH and SST• Neural Network Ideas• Back to SSH and SST • Future Pathways
2
Synoptic Ocean Prediction
• Ocean Prediction is an Initial value problem
• Features define the Initial State
• Examples of Features: Fronts, Eddies, Jets, Upwelling, Cold pools
• Time-scale of prediction: days-to-weeks
Gulf Stream Front, Eddies, Jets
Features in Western North Atlantic
• Gulf Stream• Warm Core Rings• Cold Core Rings• Southern Recirculation Gyre• Northern Recirculation Gyre• Deep Western Boundary
Current• Gangopadhyay et al., 3-part series
in 1997: Journal of Atmospheric and Oceanic Tech. (14) 1314:1365
• Maine Coastal Current• NEC Inflow• GSC Outflow• Jordan Basin Gyre• Wilkinson Basin Gyre• Georges Basin Gyre• Georges Bank Gyre• Tidal Mixing Front• Gangopadhyay et al. 2003: CSR
23 (3-4) 317-353 • Gangopadhyay and Robinson,
2002: DAO 36(2002) 201-232
Deep Sea region (GSMR) Coastal region (GOMGB)
Gulf Stream Front, Eddies, Jets
In general, a coastal current (CC), a front (SSF) and an eddy/gyre (E/G) are represented by:
CC: TM(x, η, z) =TMa(x, z)+ αM(x, z) M(η)
SSF: Tss(x, y, z) = Tsh (x, z) + (Tsl (x, z) – Tsh(x, z)) (, z)
E/G: T(r, z) = Tc (z) - [Tc (z) - Tk (z) ] {1-exp(-r/R)}
where, TM
a(x, z), Tsh (x, z) and Tc (z) are axis, shelf and core
(η) = (0 W)
(, z) = ½ + ½ tanh[(-.Z)/]
This is what is called “Feature Modeling”
Data and Feature Models
Numerical Model Initialization and
Forecast
Brown et al. (2008a, b), IEEE JOE
SST
Feature Model
July 30, 2001
FORMS Protocol
• Identify Circulation and Water mass features• Regional Synthesis -- Processes from a modeling
perspective• Synoptic Data sets -- in-situ and satellite• Regional Climatology (Background Circulation)• Multiscale Objective Analysis (Climatology +
Feature Models)• Simulation -- Nowcasting/Forecasting• Assimilation
Gulf Stream path identification -- Issues
• Historically, we have looked at SST for guidance on the North Wall
• Similar SST gradients exist elsewhere• Gulf Stream NW does not have a single
isotherm signature on the surface• Clouds• Eddies convolute the path• Large amplitude meandering to the east
often segmented11
Two Approaches
• Dynamics – Based (SST, SSH, other derived fields)
• Neural Network – Learning from the past observed paths and applying to the detection
12
13
Identifying Features
• Sea Surface Temperature (SST)• Sea Surface Height (SSH)• Sea Surface Velocity (SSUV)
14
Sea Surface Temperature
15
Sea Surface Height
16
Sea Surface Velocity
17
Extraction by Manual Operator
18
Extraction by Isoheight contouring Works better
than using SST
19
Approach 2 Neural NetworkMultilayer Perceptron (MLP)
•Type of neural network• Classification technique based on animals’ central
nervous systems•Feed forward network• Input values passed through one or more hidden
layers• Hidden layers connected between input and output
buffers•Sigmoid function applied in hidden layers•Connections between nodes in layers are
weighted•Supervised learning by backpropagation
Multilayer Perceptron visualized
The network visualized
Results visualized
• Blue dots show all points classified by network as part of GSNW
• Black line is constructed from average latitude of all blue points for a longitude
• Red line is the manual, expert-plotted line
Results visualized cont.
Conclusions• Poor results overall• Lack of variation • Indicates a possible overfitting of the
network• Overfitting results when a network
fits its output too closely to its training data
• Too many points• Possibly too low requirements for
classifying points as part of GSNW
Future plans for Neural Networks• New approach: clustering
• Used successfully in the past for feature detection• GSNW is a feature with distinct attributes
• More conducive to visual validation of results• As opposed to automated training of MLP
• Could allow for identification of entire Gulf Stream as a feature
• Takes context of points into account in a way that MLP does not
Back to Dynamics
27
28
Coming Back to SSHA Validation
0.35 m isoheoght contour
29
0.50 m isoheight contour
30
Difference between GSNW and Axis
31
Nowcast -- October 12, 2015
32
Forecast 20 October 2015
33
Future Directions with SSHA• Use the 0.5 m isoheight contour to identify a
near-axis stream path. Explore seasonality.• Use the zero-vorticity line to converge on a
finer isoheight contour (closer to the axis).• Use a parametric model (offset-curvature
dependence) to extract the North Wall• Validate and verify with concurrent SST and
SSC • Develop a mixed isoheight-zero vorticity
algorithm for eddies
34
Thank You!
35