Title: Global cyclone detection and tracking GLYDER for climate variability studies: A multisensor data pro
1Global cyclone detection and tracking (GLYDER)
for climate variability studies A multisensor
data processing approachNASA AISR Program
Ashit Talukder, Andrew Bingham, Timothy
Liu Email Ashit.Talukder_at_jpl.nasa.gov 818 354
1000 Jet Propulsion Laboratory Oct. 04, 2006
1
2Outline
- Science and Technical Objectives
- Customers/end-users
- Tech transfer integration possibilities
- Explore mission alignment
- Technical Approach
- Preliminary Results (!!)
- Plans for FY07 (Yr 1)
Hurricane Georges, Sept 24, 1998
Hurricane Floyd, Sept 14, 1999
3GLYDER Primary Science Motivation
- Earths climate system exhibits intrinsic
variability - Tropical extra-tropical cyclones important
components of Earth climate system - Key manifestations of the oceanic air-sea
interaction - Contribute to regional heat exchanges, which
affects ocean atmosphere dynamics - Complex pattern in the variability in number and
intensity of global cyclone events. - Some regions (sub tropical northeast Pacific)
experiencing increase in cyclonic frequency over
last several decades - Moderate-strong tropical cyclones have decreased
in number intensity since the 1980s (attributed
to more frequent occurrences of El Niño) - Intergovernmental Panel on Climate Change (IPCC)
has clearly identified the need to quantify the
variability in global cyclones and in particular
characterize changes in cyclone tracks - GLYDER will provide better understanding for the
reasons behind and effects of global climatic
variations via autonomous cyclone detection and
tracking
Hurricane Floyd, Sept 14, 1999
Hurricane Georges, Sept 24, 1998
4GLYDER Primary Science Objectives
- Software and IT Products to better characterize
global cyclone variability - Tools that empower JPL/NASA climate scientists to
study and quantify the spatial and temporal
variability of cyclones and their tracks - Proof of concept algorithms and software tools
for each technology component exists and has been
proven in other applications - Integrate observations from multiple remote (and
in-situ) sensors robustly and automatically - Enable yet un-achievable functionality to fuse
multiple disparate spatio-temporal measurements
for variety of earth observation needs - Extend to other maritime and terrestrial event
detection and tracking - NASAs data providers to tag metadata with
information pertaining to cyclones and enable
content-based searching - Automatically feed information to current NASA
RD projects , including GHRSST (GODAE High
Resolution Sea Surface Temperature), Earth
Science Datacasting and the Physical Oceanography
DAAC
Hurricane Dennis, August 28, 1999
Hurricane Mitch, Oct 26, 1998
5Primary Technical Objectives
- Co-register multiple remote (and in-situ)
spatio-temporal sensor observations using
multiscale techniques - Pattern recognition techniques to detect cyclones
from multiple sensors autonomously - Develop core technologies to coordinate multiple
spatially distributed in-situ and remote sensors - Track cyclones over time
- Visualize tracks of cyclones globally
- Build tools to integrate cyclone event detection
and tracking results into science (PODAAC)
software - Initiate mechanisms to tag metadata onto remote
sensing data distributed to science community
6GLYDER Customers and End Users
- Ocean/Climate Researchers need GLYDER to mine
the vast data sets and extract cyclone
information - Ocean weather data providers will use the
technology to automatically generate data
products and enable content-based searching - Ocean/Climate Researchers and Ocean/weather data
providers will use GLYDER for visualization of
global cyclone tracks - Application/operational scientists could
potentially use the technology for real-time
detection and tracking - Longer term end-user after proven off-line
operation
7Current State of Art
- Estimates of cyclone variability currently
derived from analyses of surface level pressure
(SLP) fields - Model output fields from NCEP/NCAR Reanalysis
Project based on in-situ inputs - Analyses assimilate observational and in-situ
data into a physical model to produce atmospheric
fields - Data span more than 50 years
- Measured every 6 hours on a 2.52.5 (275 km)
global grid - Accuracy of analyses is severely limited over the
oceans, - Lack of assimilated pressure and radiosonde
observations - Mean cyclone size varies from 120 440 km
Liu99 - Resolve up to a maximum of only 50 of the global
cyclones. - Satellite remote sensing provides global coverage
and greater spatial resolution, potentially
allowing detection of most/all global cyclones
Map of daily pressure observations used by ECMWF
Reanalysis forecast (6-29-02). Under sampling
in the worlds oceans especially in the Southern
Hemisphere is evident, as shown in circles.
Boxes indicate regions of interest for GLYDER
8Remote Sensors for Cyclone Detection
- Individually remote sensing datasets have
limited detection ability - Poor temporal or spatial resolution
- Loss of data due to environmental effects.
- GOES visible cloud formation
- AVHRR cloud-free surface temperature or top of
the atmosphere temperature - QuikSCAT surface wind speed and direction
- AMSR-E surface temperature
9GLYDER Primary Challenges
- Data extraction for training and validation is
non-trivial - Multiple datasets generated by different data
providers - Different data formats and file naming
conventions - Large data volumes
- Initial efforts to extract relevant data from
multiple sensors for training and performance
verification will be manually intensive and
laborious - Co-registration of multiple sensors non-trivial
- Cyclone detection
- Cyclone tracking from multiple sensors
10Multisensor Co-Registration
- Traditional interpolation and extrapolation
algorithms insufficient - Spline, NN, Bilinear etc. not suited for
non-stationary data - Initial manual mining of data tedious and
laborious - Discrete wavelet transform (DWTs) for
interpolation of each spatiotemporal
non-stationary sensory dataset - Characterize variety of smoothness function
spaces - Estimate all wavelet coefficients that describe a
signal, given a subset of the sampled data that
is noisy - parameter a in Sobolev spaces relates the rate of
decay of the DWT coefficients across scale
(frequency) low value for a denotes highly
discontinuous signals. - Problem definition
- wavelet coefficients for an M-sampled data with a
dyadic tree of size log2(M) should be estimated
based on N lt M observations. - A linear set of N equations relates the N
observations to the scale and wavelet
coefficients (for a given Sobolev smoothness a).
Since MltN, this results in an under-determined
equation, which can be solved using a
pseudo-inverse solution. - Segment signal into regions with varying
discontinuities - Each region has different a value
- Presence of discontinuities using quadratic
Bspline wavelet coefficient magnitudes
Spline interpolation fails
Segmentation of signal into 2 regions, with high
and low discontinuities using the B-spline
wavelet coefficients
11Multisensor Co-Registration
- Segment Sampled 2D data
- Using the Bspline wavelet coefficients, the 2D
data can be segmented into a library of regions
with different smoothness. - Spatial Interpolation Extend formulation of 1D
wavelet interpolation to a 2D spatial grid - For every spatial data sample, estimate the
wavelet interpolated data for a 15x15Km grid
using techniques similar to the 1D case. - The Sobolev a value is different for each
segmented region helps in better interpolation
of non-stationary signals. - Temporal Interpolation
- Obtain 2D spatially interpolated data grid for a
given sensor at a given time - Compute 1D temporal wavelet at every grid point
interpolate sensor value in time at every spatial
grid location.
Wavelet interpolation results (bottom row) using
proposed Sobolev smoothing constraint on 1D
signal using only (b) 256 of 1024 samples and
(c) 32 of 1024 samples.
(a) Correct signal (b) 256 of 1024
samples and reconstruction (c) 32 of 1024
samples
12Multisensor Cyclone Detection
- Pattern recognition fusion technique to
accurately detect cyclones - Should generalize well from limited training sets
- Nonlinear classifier for better recognition from
noisy data - Maximum discriminant feature classification
- Yields more generalized decision surfaces for
high dimensional data sets (images, multispectral
data) therefore less affected by the curse of
dimensionality - Improved cyclone detection from spatial image
datasets since it implicitly guarantees
translation-invariance in the design phase. - Proven better performance than PCA, NNs (and SVMs
in certain cases)
(a) 3 bands of a multispectral image data of M1
tank and (b) MDF fusion results with higher
signal-to-noise fusion vs. (c) Linear PCA fusion
results
13Cyclone Tracking
Tracking Using LK Features (DARPA RV2020
Program)
- Improve and adapt image feature-based tracking
techniques to multidimensional data - Prior work by JPL team on real-time tracking of
non-rigid objects for robotics - Dense optical flow (movie)
- Lucas-Kanade feature tracking
- Determine corner (Forstner) features
automatically - Track corner features using multiscale approach
- Used to determine visual odometry (camera motion)
combining stereo with temporal feature tracking - Segment camera motion from moving objects
- Track moving objects from moving camera over time
- Scale space David Lowe features
- Slower to derive, more robust
14Preliminary Cyclone Tracking Results
- QuikSCAT - polar orbiting satellite with an 1800
km wide measurement swath on the earth's surface.
- SeaWinds instrument on QuikSCAT satellite
- Specialized radar measures near-surface wind
speed direction at 25km resolution. - Launched on 19 June 1999
- Measuring winds over 90 of the ocean daily since
19 July 1999. - Temporal resolution twice per day coverage over
a given region - QuikScat wind speed 0.1o resolution, 135-190o
longitude, 5-60o latitude - Goal Track Hurrican Ioke from remote wind speed
(single sensor) data
15Preliminary Cyclone Tracking Results
- Wind speed at 0.1o resolution, 135-190o
longitude, 5-60o latitude - Goal Tracking of Hurrican Ioke from remote wind
speed (single sensor) data
16Preliminary Tracking Results
17Preliminary Tracking Results
18Schedule (FY07)
- Extract sample data subsets for previously
identified cyclones (Dec 2006) - Use for training and validation of detection and
tracking algorithms - Develop tracking algorithms to track cyclones
over time from single sensor observations (April
2007) - Extend concepts for multidimensional, multisensor
tracking - Algorithm development for co-registration of
multisensor data at different spatial and
temporal resolutions (Sept. 2007) - Interact frequently with ocean data providers and
scientists to prepare for tech. transfer in Years
1, 2, 3 - Data providers are interested in GLYDER
technology for metadata tagging - Ocean research scientists need the technology for
climate research - Application scientists can use the technology
for detecting and tracking cyclones in real-time)
19Personnel/Workforce
- Project commences Oct 1, 2006
- Technology Development
- Dr. A. Talukder (PI) - Ashit.Talukder_at_jpl.nasa.go
v - Dr. Anand Panangadan (Post-doc)
- NASA NPP Post-Doctoral Scholar
- Science Validation and Tech Transfer
- Dr. Andrew Bingham JPL
- Dr. Timothy Liu JPL