[Met_help] [rt.rap.ucar.edu #59713] History for objects in MODE
John Halley Gotway via RT
met_help at ucar.edu
Tue Mar 26 09:57:28 MDT 2013
----------------------------------------------------------------
Initial Request
----------------------------------------------------------------
I am now working on some grid data using MODE of METV4.0. I understand the logic of generating the objects, but I would like to know how are the attributes of the objects calculated. For example (see attached), does the area of cluster object in observation field include the "white" area since the affecting area (the "L" shape) is not that large. If yes, is there any way to tune the config such that we can refine the cluster objects to fit the actual shape.
Thanks.
Johnny
----------------------------------------------------------------
Complete Ticket History
----------------------------------------------------------------
Subject: Re: [rt.rap.ucar.edu #59713] objects in MODE
From: John Halley Gotway
Time: Fri Jan 04 12:27:59 2013
Johnny,
I apologize for the delay in getting back to you. Things were slowed
down over the holidays.
Thank you for sending the postscript file to illustrate you question.
The answer is no, the "white" area is not included in the area of the
cluster object. What's likely causing the confusion is the
black outline that's drawn around the L-shaped object. We call that
black outline the convex hull, which is used in the computation of
some of the object attributes, but not area. Often cluster
objects consist of more than one simple object. Drawing the convex
hull around that group, along with the color of the objects, helps
illustrate the grouping.
For each combination of simple forecast and simple observation object,
several object pair attributes are computed. The values are scaled to
a [0,1] range using the interest maps defined in the MODE
configuration file. And a weighted average of the object pair scores
is computed as the "total interest" for that pair.
Those object pair attribute weights are listed on the first page of
the postscript image. They are as follows:
- "Centroid" for the distance between the forecast and observation
object centroids.
- "Boundary" for the minimum distance between the forecast and
observation objects.
- "Convex Hull" for the minimum distance between the forecast and
observation object convex hulls.
- "Angle" for the difference between the forecast and observation
object's orientation angles.
- "Area" for the ratio of the forecast and observation object areas.
1.0 means they're exactly the same size.
- "Intersection" for the ratio of their intersection area to their
union area. 1.0 means they're exactly the same.
- "Complexity" for the ratio of their individual complexity values,
where complexity is based on the ratio of the object area to the
convex hull area.
- "Intensity" for the ratio of the selected intensity percentile of
interest. For example, if the intensity percentile is set to 50, or
the median, this is the ratio of the medians intensity of the
values within the forecast object to that of the observation object.
The object pair attributes are combined into a single total interest
value for the pair. Pair with a total interest above the threshold
(0.7 in your example) are matched. Those with total interest
less than that are not matched.
So there's a lot of knobs to tune. Generally, we use attributes about
the location, such as boundary distance, centroid distance, and
intersection area to drive how the matching is performed. In
your example, that single match is a pretty obvious one. But as the
number of objects increase, it's more important to tune the weights to
get the types of matches you deem appropriate.
Hope that helps clarify.
John Halley Gotway
met_help at ucar.edu
On 12/27/2012 03:40 AM, Tam Johnny via RT wrote:
>
> Thu Dec 27 03:40:02 2012: Request 59713 was acted upon.
> Transaction: Ticket created by johnnytam_hko at yahoo.com.hk
> Queue: met_help
> Subject: objects in MODE
> Owner: Nobody
> Requestors: johnnytam_hko at yahoo.com.hk
> Status: new
> Ticket <URL:
https://rt.rap.ucar.edu/rt/Ticket/Display.html?id=59713 >
>
>
> I am now working on some grid data using MODE of METV4.0. I
understand the logic of generating the objects, but I would like to
know how are the attributes of the objects calculated. For example
(see attached), does the area of cluster object in observation field
include the "white" area since the affecting area (the "L" shape) is
not that large. If yes, is there any way to tune the config such that
we can refine the cluster objects to fit the actual shape.
>
> Thanks.
> Johnny
>
------------------------------------------------
Subject: objects in MODE
From: Tam Johnny
Time: Mon Feb 25 00:39:21 2013
Dear John,
Thank you for your detailed explanation. I would also like to know
more about the calculation of total interest. is it the value 1.0
meaning a perfect match? I have a case (see attached) that the total
interest is 1.0, but there is a slightly difference between the obs
and the forecast. is there any wrong for my weighting setting?
Regards,
Johnny
Johnny,
I apologize for the delay in getting back to you. Things were slowed
down over the holidays.
Thank you for sending the postscript file to illustrate you question.
The answer is no, the "white" area is not included in the area of the
cluster object. What's likely causing the confusion is the
black outline that's drawn around the L-shaped object. We call that
black outline the convex hull, which is used in the computation of
some of the object attributes, but not area. Often cluster
objects consist of more than one simple object. Drawing the convex
hull around that group, along with the color of the objects, helps
illustrate the grouping.
For each combination of simple forecast and simple observation object,
several object pair attributes are computed. The values are scaled to
a [0,1] range using the interest maps defined in the MODE
configuration file. And a weighted average of the object pair scores
is computed as the "total interest" for that pair.
Those object pair attribute weights are listed on the first page of
the postscript image. They are as follows:
- "Centroid" for the distance between the forecast and observation
object centroids.
- "Boundary" for the minimum distance between the forecast and
observation objects.
- "Convex Hull" for the minimum distance between the forecast and
observation object convex hulls.
- "Angle" for the difference between the forecast and observation
object's orientation angles.
- "Area" for the ratio of the forecast and observation object areas.
1.0 means they're exactly the same size.
- "Intersection" for the ratio of their intersection area to their
union area. 1.0 means they're exactly the same.
- "Complexity" for the ratio of their individual complexity values,
where complexity is based on the ratio of the object area to the
convex hull area.
- "Intensity" for the ratio of the selected intensity percentile of
interest. For example, if the intensity percentile is set to 50, or
the median, this is the ratio of the medians intensity of the
values within the forecast object to that of the observation object.
The object pair attributes are combined into a single total interest
value for the pair. Pair with a total interest above the threshold
(0.7 in your example) are matched. Those with total interest
less than that are not matched.
So there's a lot of knobs to tune. Generally, we use attributes about
the location, such as boundary distance, centroid distance, and
intersection area to drive how the matching is performed. In
your example, that single match is a pretty obvious one. But as the
number of objects increase, it's more important to tune the weights to
get the types of matches you deem appropriate.
Hope that helps clarify.
John Halley Gotway
met_help at ucar.edu
On 12/27/2012 03:40 AM, Tam Johnny via RT wrote:
>
> Thu Dec 27 03:40:02 2012: Request 59713 was acted upon.
> Transaction: Ticket created by johnnytam_hko at yahoo.com.hk
> Queue: met_help
> Subject: objects in MODE
> Owner: Nobody
> Requestors: johnnytam_hko at yahoo.com.hk
> Status: new
> Ticket <URL:
https://rt.rap.ucar.edu/rt/Ticket/Display.html?id=59713 >
>
>
> I am now working on some grid data using MODE of METV4.0. I
understand the logic of generating the objects, but I would like to
know how are the attributes of the objects calculated. For example
(see attached), does the area of cluster object in observation field
include the "white" area since the affecting area (the "L" shape) is
not that large. If yes, is there any way to tune the config such that
we can refine the cluster objects to fit the actual shape.
>
> Thanks.
> Johnny
>
------------------------------------------------
Subject: Re: [rt.rap.ucar.edu #59713] objects in MODE
From: John Halley Gotway
Time: Mon Feb 25 08:41:10 2013
Johnny,
Good question - how can you get a total interest of 1, when the two
objects are not identical?
Please take a look at section 6.2.3 of the MET User's Guide, titled
"Fuzzy Logic":
http://www.dtcenter.org/met/users/docs/users_guide/MET_Users_Guide_v4.0.1.pdf
There are 8 pair attributes that go into the computation of that total
interest value. For each of those 8, there is a weight and an
interest function defined in the configuration file. (Note that
there is actually a 3rd term, called confidence, but it doesn't
warrant discussion here.) The weights are defined in the "weight"
section while the interest functions can be found in the
"interest_function" section of the config file. The interest function
is used to map something with units (like a distance in grid squares)
to a unit-less value between 0 and 1. The total interest
is just a weighted average of those 8 values between 0 and 1.
Each interest function is just a piecewise-linear function defined by
the points specified in the config file. Notice that the distance-
related interest functions (centroid_dist, boundary_dist, and
convex_hull_dist) are defined relative to the "grid_res" parameter,
defined at the top of the config file. Let's take a look at the
centroid-distance interest function:
centroid_dist = (
( 0.0, 1.0 )
( 60.0/grid_res, 1.0 )
( 600.0/grid_res, 0.0 )
);
Since the grid_res is set to 4 in the default config file, the 3
points in this piece-wise linear function are: (0, 1), (15, 1), and
(150, 0). So any two objects whose centroids are between 0 and 15
grid squares away from each other are assigned a value of 1.0 for the
centroid distance attribute.
In the PostScript file you sent, I see that you have non-zero weights
for centroid distance, boundary distance, angle difference, area
ratio, and intersection/union area. It just must be the case
that those 5 parameters were "close enough" to warrant a total
interest value of 1.
Here are some things you could try:
- Ensure that you have the "grid_res" parameter set to the
approximate size of each grid box in kilometers.
- Rerun your case setting the verbosity level to 5 (-v 5). When you
do, you'll see detailed information about each of the 8 pair
attributes that go into the total interest.
Hope that helps.
Thanks,
John
On 02/25/2013 12:39 AM, Tam Johnny via RT wrote:
>
> <URL: https://rt.rap.ucar.edu/rt/Ticket/Display.html?id=59713 >
>
> Dear John,
>
> Thank you for your detailed explanation. I would also like to know
more about the calculation of total interest. is it the value 1.0
meaning a perfect match? I have a case (see attached) that the total
interest is 1.0, but there is a slightly difference between the obs
and the forecast. is there any wrong for my weighting setting?
>
> Regards,
> Johnny
>
>
>
> Johnny,
>
> I apologize for the delay in getting back to you. Things were
slowed down over the holidays.
>
> Thank you for sending the postscript file to illustrate you
question. The answer is no, the "white" area is not included in the
area of the cluster object. What's likely causing the confusion is
the
> black outline that's drawn around the L-shaped object. We call that
black outline the convex hull, which is used in the computation of
some of the object attributes, but not area. Often cluster
> objects consist of more than one simple object. Drawing the convex
hull around that group, along with the color of the objects, helps
illustrate the grouping.
>
> For each combination of simple forecast and simple observation
object, several object pair attributes are computed. The values are
scaled to a [0,1] range using the interest maps defined in the MODE
> configuration file. And a weighted average of the object pair
scores is computed as the "total interest" for that pair.
>
> Those object pair attribute weights are listed on the first page of
the postscript image. They are as follows:
> - "Centroid" for the distance between the forecast and observation
object centroids.
> - "Boundary" for the minimum distance between the forecast and
observation objects.
> - "Convex Hull" for the minimum distance between the forecast and
observation object convex hulls.
> - "Angle" for the difference between the forecast and observation
object's orientation angles.
> - "Area" for the ratio of the forecast and observation object areas.
1.0 means they're exactly the same size.
> - "Intersection" for the ratio of their intersection area to their
union area. 1.0 means they're exactly the same.
> - "Complexity" for the ratio of their individual complexity values,
where complexity is based on the ratio of the object area to the
convex hull area.
> - "Intensity" for the ratio of the selected intensity percentile of
interest. For example, if the intensity percentile is set to 50, or
the median, this is the ratio of the medians intensity of the
> values within the forecast object to that of the observation object.
>
> The object pair attributes are combined into a single total interest
value for the pair. Pair with a total interest above the threshold
(0.7 in your example) are matched. Those with total interest
> less than that are not matched.
>
> So there's a lot of knobs to tune. Generally, we use attributes
about the location, such as boundary distance, centroid distance, and
intersection area to drive how the matching is performed. In
> your example, that single match is a pretty obvious one. But as the
number of objects increase, it's more important to tune the weights to
get the types of matches you deem appropriate.
>
> Hope that helps clarify.
>
> John Halley Gotway
> met_help at ucar.edu
>
>
> On 12/27/2012 03:40 AM, Tam Johnny via RT wrote:
>>
>> Thu Dec 27 03:40:02 2012: Request 59713 was acted upon.
>> Transaction: Ticket created by johnnytam_hko at yahoo.com.hk
>> Queue: met_help
>> Subject: objects in MODE
>> Owner: Nobody
>> Requestors: johnnytam_hko at yahoo.com.hk
>> Status: new
>> Ticket <URL:
https://rt.rap.ucar.edu/rt/Ticket/Display.html?id=59713 >
>>
>>
>> I am now working on some grid data using MODE of METV4.0. I
understand the logic of generating the objects, but I would like to
know how are the attributes of the objects calculated. For example
(see attached), does the area of cluster object in observation field
include the "white" area since the affecting area (the "L" shape) is
not that large. If yes, is there any way to tune the config such that
we can refine the cluster objects to fit the actual shape.
>>
>> Thanks.
>> Johnny
>>
>
>
------------------------------------------------
More information about the Met_help
mailing list