Automated Cyclone Discovery and Tracking using Knowledge Sharing in Multiple Heterogeneous Satellite Data

1
Automated Cyclone Discovery and Tracking using
Knowledge Sharing in Multiple Heterogeneous
Satellite Data
Authors Shen-Shyang Ho Ashit Talukder Jet
Propulsion Laboratory California Institute of
Technology

• Group 3
• Karen Simpson
• Paul Fomenky
• Roman Sizov
• Sameh Ebeid

Assignment 1 02/22/2010
2
Outline

• Introduction
• Previous Work
• Data Description
• Issues and Challenges
• Heterogeneous Remote Satellite-Based Detection
  and Tracking Approach
• Experimental Results
• Lessons Learned and Conclusions

3
What is Cyclone
Introduction

• An area of closed, circular fluid motion rotating
  in the same direction as the Earth
• Low pressure areas, their center is the lowest
  atmospheric pressure in the region

4
Surface-based Types
Introduction

• Polar cyclone
• Polar low
• Extra-tropical
• Sub-tropical
• Tropical
• Mesoscale

5
Extra-tropical
Introduction

• Synoptic scale low pressure weather system that
  has neither tropical nor polar characteristics
• Often described as depressions or lows by weather
  forecasters

6
Tropical
Introduction

• Storm characterized by a low pressure center and
  numerous thunderstorms that produce strong winds
  and flooding rain
• Referred to by other names such as hurricane,
  typhoon, tropical storm
• Develop over large bodies of warm water, and lose
  strength if they move over land

7
Tropical
Introduction

• An average 86 tropical cyclones of tropical storm
  intensity form annually worldwide, 47 reaching
  hurricane/typhoon strength, and 20 becoming
  intense tropical cyclones

8
(No Transcript)
9
Cyclone detection and tracking
Introduction

• The tropical prediction center / National
  Hurricane Center (TPC/NHC) use conventional
  surface and upper-air observations and
  reconnaissance aircraft report
• In recent years, some studies have used satellite
  images that are manually retrieved and analyzed
  to improve the accuracy of cyclone tracking

10
Cyclone detection and tracking
Introduction

• A new automated global cyclone discovery and
  tracking approach on a truly global basis using
  near real-time (NRT) and historical sensor data
  from multiple satellite
• This implementation employs two types of
  satellite sensor measurements
• QuikSCAT wind satellite data
• Merged precipitation data using TRMM and other
  satellites

11
Cyclone detection and tracking
Introduction

• Challenges pertaining to mining data from
  orbiting satellites
• Each orbiting satellite cannot monitor a region
  continuously and the measurements are
  instantaneous

12
Cyclone detection and tracking
Introduction

• Challenges pertaining to mining data from
  orbiting satellites
• Each orbiting satellite cannot monitor a region
  continuously and the measurements are
  instantaneous

Can minimize their effects by using data from
multiple satellite
13
Cyclone detection and tracking
Introduction

• Challenges pertaining to mining data from
  orbiting satellites
• Each orbiting satellite cannot monitor a region
  continuously and the measurements are
  instantaneous
• Different satellites provide different
  measurements
• Different satellites sensors acquire measurements
  at different spatial and temporal resolution

14
Cyclone detection and tracking
Introduction
These problems make mining heterogeneous data
from multiple orbiting satellites extremely
challenging and remains as a now primarily an
unsolved problem

• Challenges pertaining to mining data from
  orbiting satellites
• Each orbiting satellite cannot monitor a region
  continuously and the measurements are
  instantaneous
• Different satellites provide different
  measurements
• Different satellites sensors acquire measurements
  at different spatial and temporal resolution

15
Cyclone detection and tracking
Introduction

• Challenges related to the problem of detection
  and tracking of cyclones
• Cyclone events are dynamic in nature
• There is lack of annotated negative (non-cyclone)
  examples by experts
• A single satellite sensor may miss a cyclone
  event due to a pre-defined orbiting trajectory

16
Outline

• Introduction
• Previous Work
• Data Description
• Issues and Challenges
• Heterogeneous Remote Satellite-Based Detection
  and Tracking Approach
• Experimental Results
• Lessons Learned and Conclusions

17
Previous work
Previous work

• No solution currently exists that uses
  heterogeneous sensor measurement to automatically
  detect and track cyclones
• The current solutions involve human interference
  and decision

18
Outline

• Introduction
• Previous Work
• Data Description
• Issues and Challenges
• Heterogeneous Remote Satellite-Based Detection
  and Tracking Approach
• Experimental Results
• Lessons Learned and Conclusions

19
QuikSCAT Wind Data
Data description

• The QuikSCAT (Quick Scatterometer) mission
  provide important high quality ocean wind data
  set
• Recent research showed QuikSCAT data is useful
  for early detection of tropical cyclones

20
Precipitation Data from TRMM satellite
Data description

• The Tropical Rainfall Measurement Mission (TRMM)
  is a joint mission between NASA and JAXA designed
  to monitor and study tropical rainfall
• The (Level) 3b-42 TRMM data product used in this
  paper is produced using a combined instrument
  rain calibration algorithm

21
Outline

• Introduction
• Previous Work
• Data Description
• Issues and Challenges
• Heterogeneous Remote Satellite-Based Detection
  and Tracking Approach
• Experimental Results and Conclusions

22
Issues and Challenges

• Main issues and challenges
• Non-Continuous Region Monitoring
• Event Occlusion
• Varying Temporal and Spatial Resolution
• Lack of Annotated Negative Examples

23
Main issues and challenges

• Satellite measurements are instantaneous hence,
  satellites cannot measure sustained winds.
  Remember, a leading characteristic of cyclones is
  sustained winds
• TRMM 3B42 data is known to underestimate
  rainfall, which might lead to false negatives

24
Non-Continuous Region Monitoring Problem

• Geostationary Operational Environmental
  Satellites (GOES) monitor specific area at all
  times, helping identify sustained winds etc.
  Unfortunately, most countries do not have these.
• Because QuikSCAT and TRMM are motile, this
  monitoring is lost. This results in invisible
  swaths.

25
Non-Continuous Region Monitoring Problem
Evidence
print
26
Non-Continuous Region Monitoring Operational
Weather Satellite System

• Satellite systems consist of two types
• Geostationary Operational Environmental
  Satellites are static and throw light on current
  and short term weather trends.
• Orbiting satellites like QuikSCAT and TRMM help
  with longer term forecasting.

27
Non-Continuous Region Monitoring Solution

• Usage of multiple satellites produces a higher
  temporal density hence helping alleviate the
  problem.
• A group of complementary satellites can make this
  problem almost insignificant.

28
Event Occlusion - Problem

• Satellite swath can partially (or worst case,
  totally) miss events of interest.
• Though in continuous orbit, event can be gone by
  time satellite comes back.

29
Event Occlusion Problem Evidence 1
print

• QuikSCAT showing only a small part of event of
  interest.
• Hurricane Dean Aug 17th 2007, 0900

30
Event Occlusion Problem Evidence 2
print

• Next QuikSCAT swath shows a bit more.
• Hurricane Dean Aug 17th 2007, 1041

31
Event Occlusion Problem Evidence 3
print

• Another QuikSCAT swath shows much more, but
  missing eye of storm.
• Hurricane Dean Aug 17th 2007, 2310

32
Event Occlusion Problem Evidence 4
print

• QuikSCAT swath from previous day showed more!
• Hurricane Dean Aug 16hth 2007, 2156

33
Event Occlusion Solution

• Clearly, multiple orbits of the same satellite
  can produce more information on the event being
  examined.
• Also, as in continuity monitoring issue, numerous
  satellites working together are less likely to
  miss important events.

34
Varying Temporal and Spatial Resolution Problem

• Different aspects influence the temporal
  resolution of measurements
• Satellite orbit time (QuikSCAT 101 minutes, TRMM
  92.5mins)
• Swath width of measuring instrument (SeaWinds on
  QuikSCAT 1800km PR, TMI and VIRS on TRMM 247km,
  878km, 873km respectively)
• Geographic coverage (QuikSCAT global TRMM
  50N to 50S)

35
Varying Temporal and Spatial Resolution Problem
Contd

• Spatial resolution depends on
• Sensor instruments (PR, TMI and VIRS on TRMM
  5.1km, 5.0km, 2.4km respectively)
• Satellite orbital altitude ((TRMM Pre-boost
  (350km) (TMI) 4.4km to 5.1km (Post-boost (403
  km))
• Processing algorithm (operational QuikSCAT data
  has spatial resolutions of 12.5km and 25km )

36
Varying Temporal and Spatial Resolution Problem
Contd 2

• In addition to inter satellite differences, there
  are some intra satellite tempo-spatial
  differences.
• TRMM Level 3 data has lower temporal resolution
  than levels 1 and 2.

37
Varying Temporal and Spatial Resolution Solution

• On TRMM, mine areas QuikSCAT showed events of
  interest on.
• Also, because of different swath sizes, latitudes
  and longitudes were used to identify locations.
• Temporal tracking done on TRMM as temporal
  resolution higher than in QuikSCAT.

38
Lack of Annotated Negative examples - Problem

• Scientists have not clearly shown what a
  non-event is despite the large archives of
  events.

39
Lack of Annotated Negative examples - Solution

• Random non-event days were monitored and fed to
  system as examples of non event.

40

• Introduction
• Previous Work
• Data Description
• Issues and Challenges
• Heterogeneous Remote Satellite-Based Detection
  and Tracking Approach
• Experimental Results and Conclusions

41
Heterogeneous Remote Satellite-Based Detection
and Tracking Approach
Heterogeneous Remote Satellite-Based Detection
and Tracking Approach

• QuikSCAT Feature Selection
• Ensemble Classifier for Cyclone Detection
• Knowledge Sharing between TRMM and QuikSCAT data
  for Cyclone Tracking

42
QuikSCAT Feature Selection
Heterogeneous Remote Satellite-Based Detection
and Tracking Approach

• Features that characterize and identify a cyclone
  are selected from QuikSCAT satellite data
• The QuikSCAT Level 2B data that consist of ocean
  wind vector information are utilized
• The Level 2B data are grouped by rows of wind
  vector cells (WVC) which are squares of dimension
  25 km or 12.5 km

43
QuikSCAT Feature Selection (contd)
Heterogeneous Remote Satellite-Based Detection
and Tracking Approach

• 1624 WVC rows at 25 km or 3248WVC rows at 12.5
  are required to cover the earth circumference
• Out of 25 fields in the data structure for the
  Level 2B data we are interested only in latitude,
  longitude, wind speed(WS) and wind direction (WD)

44
QuikSCAT Feature Selection (contd)
Heterogeneous Remote Satellite-Based Detection
and Tracking Approach

• Table 1. The fields of interest from Level 2B
  data structure

Field Unit Minimum Maximum
WVC latitude Deg -90.00 90.00
WVC longitude Deg E 0.00 359.99
Selected speed m/s 0.00 50.00
Selected direction Deg from North 0.00 359.99
45
QuikSCAT Feature Selection (contd)
Heterogeneous Remote Satellite-Based Detection
and Tracking Approach

• The Level 2B data needs to be interpolated on a
  uniformly gridded surface due to the
  non-uniformity in the measurements taken by the
  QuikSCAT satellite on a spherical surface
• The nearest neighbor rule is used for this
  pre-processing procedure for both wind speed (WS)
  and wind direction (WD)

46
QuikSCAT Feature Selection (contd)
Heterogeneous Remote Satellite-Based Detection
and Tracking Approach

• Histograms are constructed to estimate
  probability density of the wind speed (WS) and
  wind direction (WD) within a predefined bounding
  box extracted from a QuikSCAT image

47
QuikSCAT Feature Selection (contd)
Heterogeneous Remote Satellite-Based Detection
and Tracking Approach

• WS(i,j),WD(i,j) wind speed and wind direction
  at location (i,j)
• DSR(i,j) the direction to speed ratio at (i,j)

48
QuikSCAT Feature Selection (contd)
Heterogeneous Remote Satellite-Based Detection
and Tracking Approach

• When there is a strong wind with wind
  circulation, the DSR at a WVC will be small
• DSR histogram will have a skewed distribution
  towards the smaller value
• When there is weak or no wind with no
  circulation, DSR histogram does not have the
  skewed characteristics

49
QuikSCAT Feature Selection (contd)
Heterogeneous Remote Satellite-Based Detection
and Tracking Approach

• When a region contains a cyclone, the WS
  histogram shows a density estimate skewed towards
  the larger values and WD histogram shows a near
  uniform distribution
• A cyclone is defined as a warm-core non-frontal
  synoptic-scale system, with organized deep
  convection and a closed surface wind circulation
  about a well-defined center

50
QuikSCAT Feature Selection (contd)
Heterogeneous Remote Satellite-Based Detection
and Tracking Approach

• To discriminate between cyclone and non-cyclone
  events based on the circulation property two
  additional features are used
• (1) a measure of relative strength of the
  dominant wind direction (DOWD)
• (2) the relative wind vorticity (RWV)

51
QuikSCAT Feature Selection (contd)
Heterogeneous Remote Satellite-Based Detection
and Tracking Approach

• u(i,j) and v(i,j) are the u-v components of the
  wind direction WD(i,j) at location (i,j) with
• 1i m and 1jn
• The (mn)-by-2 matrices M are constructed as
  follows

52
QuikSCAT Feature Selection (contd)
Heterogeneous Remote Satellite-Based Detection
and Tracking Approach

• If ?1 and ?2 are the eigenvalues of matrix M such
  that ?1 lt ?2, then the eigenvalue ratio of a
  bounding box B of dimension m by n is
• ERB is used to quantify the relative strength of
  the dominant wind direction (DOWD) within B

53
QuikSCAT Feature Selection (contd)
Heterogeneous Remote Satellite-Based Detection
and Tracking Approach

• If there is a circulation (i.e. a cyclone in B),
  ERB will be near to 1
• If the wind is unidirectional (no storm or
  cyclone in B), ?2 will be much greater than ?1,
  and as a result ERB is much larger

54
QuikSCAT Feature Selection (contd)
Heterogeneous Remote Satellite-Based Detection
and Tracking Approach

• The relative wind vorticity (RWV) at location
  (i,j) is calculated by the formula
• where u and v are the two wind vector components
  in the west-east and south-north directions, and
  d is the spatial distance between two adjacent
  QuikSCAT measurements in a uniformly gridded data

• ?z or ? vertical component of relative vorticity

?z or ? vertical component of
relative vorticity
55
Ensemble Classifier for Cyclone Detection
Heterogeneous Remote Satellite-Based Detection
and Tracking Approach

• Ensemble methods are learning algorithms that
  make predictions on observations based on a
  majority or weighted vote from a set of
  classifiers or predictors

56
Ensemble Classifier for Cyclone Detection (contd)
Heterogeneous Remote Satellite-Based Detection
and Tracking Approach

• The ensemble classifier is built to identify
  cyclones in QuikSCAT images
• The TRMM precipitation data are not used in the
  ensemble because
• It has a weak discriminating power heavy
  rainfall does not imply existence of cyclone
• It is very unlikely that one has QuikSCAT and
  TRMM data concurrently

57
Ensemble Classifier for Cyclone Detection (contd)
Heterogeneous Remote Satellite-Based Detection
and Tracking Approach

• Regions in a QuikSCAT image likely to contain a
  cyclone are localized based on wind speed
• Regions that have areas less than some threshold
  are removed
• Five classifiers based on features extracted from
  the QuikSCAT training data are constructed to
  identify the cyclones
• Two classifiers are thresholding classifier based
  on the DOWD and RWV features, and the other three
  are support vector machine (SVM) that use
  histogram features for WS, WD and DSR
• The classification decision is based on majority
  vote among the five classifiers

• Figure 5. Ensemble Classifier (Cyclone Discovery
  Module)

58
Knowledge Sharing between TRMM and QuikSCAT data
for Cyclone Tracking
Heterogeneous Remote Satellite-Based Detection
and Tracking Approach

• The multi-sensor knowledge-sharing solution is
  based on the strength of each remote sensor type
• QuikSCAT has excellent information for cyclone
  detection but lack sufficient temporal resolution
  (each pass-through is repeated only every 12
  hours)
• TRMM has excellent temporal resolution of 3
  hours, but lacks good discriminative ability for
  accurate cyclone detection
• Therefore, QuikSCAT data are used for cyclone
  detection, and TRMM data for tracking based on
  knowledge obtained from the ensemble classifier
  using QuikSCAT features

59
Knowledge Sharing between TRMM and QuikSCAT data
for Cyclone Tracking (contd)
Heterogeneous Remote Satellite-Based Detection
and Tracking Approach

• QuikSCAT data are retrieved, and are input into
  the cylone discovery module to locate or identify
  possible cyclones
• The cyclone location is used to predict the
  likely regions to contain a cyclone at the next
  incoming data stream retrieved using a linear
  Kalman filter predictor, which is important
  because TRMM precipitation data are not a
  definitive indicator of cyclones
• A cyclone localized by applying a threshold to
  the TRMM precipitation rate measurement (T6 0)
• After a cyclone is located the Kalman filter
  measurement update or correction is applied to
  obtain an estimate of the new state vector or the
  predicted location of the cyclone in the next
  TRMM (or QuikSCAT) observation cycle

• Figure 6. Knowledge sharing between TRMM and
  QuikSCAT data for Cyclone Tracking

60
Knowledge Sharing between TRMM and QuikSCAT data
for Cyclone Tracking (contd)
Heterogeneous Remote Satellite-Based Detection
and Tracking Approach

• A cyclone is a dynamic event and its size evolves
  rapidly over time, and therefore modeling and
  predicting only the cyclone center in space over
  time would be grossly inadequate
• Thus, the maximum and the minimum
  latitude/longitude of the bounding box spanned by
  the cyclone is used based on the hypothesis that
  the cyclone evolves linearly in space over time
• The estimated bounding box was expanded (or
  contracted) based on the estimated Kalman error
  covariance to define a search region for the
  cyclone in the TRMM image
• This modeling significantly improves the quality
  of knowledge sharing between heterogeneous
  satellites compare to the model that uses only
  the center coordinates of the cyclone

61
Outline

• Introduction
• Previous Work
• Data Description
• Issues and Challenges
• Heterogeneous Remote Satellite-Based Detection
  and Tracking Approach
• Experimental Results and Conclusions

62
Training Set and Test Data
Experimental Results

• Training Set
• 191 QuikSCAT images of cyclones occurring in
  North Atlantic Ocean in 2003
• 1833 negative examples (unlabeled examples from
  four days in 2003 that no tropical cyclone)
• Test Set
• 54 cyclone events in North Atlantic Ocean in 2006
• 1822 non-cyclone events

63
Classification Performance
Experimental Results

• Step 1 Determine thresholds for DOWD (Dominant
  Wind Direction) and RWV (Dominant Wind Vorticity)
  features from test set results

64
Performance of DOWD classifier
Experimental Results
0.80
Positive
0.59
0.38
1.958
65
Performance of RWV Classifier
Experimental Results
0.89
0.85
0.80
1.51
66
Experimental Results
Classification Performance

• Step 1 Determine thresholds for DOWD (Dominant
  Wind Direction) and RWV (Dominant Wind
  Vorticity) features from test set results
• Step 2 Analyze performance of different
  classifier ensembles

67
Different Classifier Ensembles
Experimental Results

• RWV
• DOWD
• SVM ensemble
• CDM
• SVM RWV ensemble
• SVM DOWD ensemble
• CIS (Ho and Talukder, 2008)

68
Experimental Results
69
ROC Curve
Experimental Results
(Receiver Operating Characteristics)

• RWV is a more robust feature than DOWD in
  discriminating cyclone and non-cyclone events

70
Classifier Performance
Experimental Results

• Step 1 Determine thresholds for DOWD (Dominant
  Wind Direction) and RWV (Dominant Wind
  Vorticity) features from test set results
• Step 2 Analyze performance of classifiers
• Step 3 Use CDM to track an isolated hurricane
  event (Hurricane Isabel, 2003) using QuikSCAT and
  TRMM data

71
Tracking Hurricane Isabel
Experimental Results
72
Experimental Results
Tracking Hurricane Isabel
73
Experimental Results
Tracking Hurricane Isabel
74
Experimental Results
Tracking Hurricane Isabel
75
Experimental Results
Tracking Hurricane Isabel
76
Experimental Results
Experimental Results
Tracking Hurricane Isabel
77
Experimental Results
Tracking Hurricane Isabel
78
Experimental Results
Tracking Hurricane Isabel
79
Experimental Results
Tracking Hurricane Isabel
80
Experimental Results
Tracking Hurricane Isabel
81
Experimental Results
Tracking Hurricane Isabel
82
Experimental Results
Tracking Hurricane Isabel
83
Experimental Results
Tracking Hurricane Isabel
84
Experimental Results
Tracking Hurricane Isabel
85
Experimental Results
Tracking Hurricane Isabel
86
Experimental Results
Tracking Hurricane Isabel
87
Conclusions

• Conventional methods that utilize human resources
  cannot handle massive, unlabeled high-dimensional
  heterogeneous data
• This method provides an efficient solution to
  track cyclonic events which combines information
  from multiple satellites
• The threshold values depend on the desired
  accuracy, as well as the desired rate of true
  positives and true negatives

88
Questions?