IRI-2005 Notes and Comments:

Slide 1: (Title)

Slide 2: (overview)

Slide 3: (intro)

Slide 4: (data)
+ apolgize for hard to read platform names

Slide 5: (hwind)
+ Remember that first slide is 'blank'
+ Note that in the example there are ~15 platforms ~3hrs of data
+ Sample is from Hurricane Ivan at landfall
+ Mention how there are areas of hi and low coverage
+ Describe how the wind barbs work (winds heading to target)
+ Briefly mention why "flagged" winds are flagged
+ make note to the fact that the current manual QC is 'subjective'
	
Slide 6: (clustering)
+ Mention how and why Clustering algorithms were noted to be useful
+ spatial clustering useful to dectect noise and outliers
+ weeding out noise good, outliers bad
+ DBSCAN was selected basically due to its relative simplicity,
  its use of 'density' and its great efficiency
+ neighbors: density reachable
+ 2 types of points:
+ core points-> those with a minimum # of neighbors
+ border points -> those without a minimum # of neighbors but neighbor
  to a core point
+ minimum # neighbors (Eps) 4, max distance based on attributes
+ key issue in all clustering... what defines a neighbor?
  = "how to define a neighbor within a hurricane?"

Slide 7: (coordinate systems)
+ trying to determine distance function for hurricanes lead to
  researching coordinate systems.
+ discuss "idealized" cylindrical shape of hurricane

Slide 8: (plots)
+ minimum distance set to 26km based on QSCAT resolution
+ still leaves much of data without neighbors
+ when cylindrical, almost no neighbor-less points
+ this reduces the number of points considered outliers
+ describe the 3 schemes used

Slide 9: (max diffs)

Slide 10: (Ivan)
+ point out radii numbers in top left corner
+ comment on colors near center for No QC and Cartesian

Slide 11: (Frances)
+ similarity between No QC and Cartesian
+ point out completly wrong max wind found in NW quad

Slide 12: (Jeanne US)
+ similiar to Ivan, poor near center for No QC and Cartesian
+ point out how the radial and polar do considerably better but still
  have wrong max wind

Slide 13: (Jeanne Hispanola)
+ all do poor
+ poor data coverage and lack of cylindrical shape

Slide 14: (results)
+ shows radii values

Slide 15: (errors)
+ mean %error
+ overall "radial" did best, Cartesian performed as well as No QC

Slide 16:
+ every storm is different
+ therefore even though radial out perform 10-14%, it was completely
  out to lunch on the last Jeanne case