Open alexcpn opened 1 month ago
The NLCD common data is based on a retiling operation of the original NLCD data snapshot of 2011 (re-released in 2014), using a set of provided scripts found in https://github.com/Wireless-Innovation-Forum/Common-Data/tree/master/scripts
Obviously if you compare the data with latest data from 2020+, then you will have differences here and there.
@sergebdt is the expected results calculated using the NLCD common data by WINNF or by third party using the latest ones. Wanted to confirm if this was indeed the reason for the differences with my calculated value and test output
I guess that any reference test output should use a common database that everyone can agree on - ie. the NLCD common data. Check with the AFC people for confirmation.
In SAS, that was the case and every production system out there reproduces internally that common NLCD database in their live system (and also using similar data read routines) so as to avoid small discrepancies (such as reading at the border, etc..). If someone use another database, there is strong risk of failing the compliance tests.
Remember also that in the common NLCD database, the format used after retiling is regular grids in degrees space (1 arcsec separation), which is different from the original NLCD source file used to construct this database.
I am seeing here also similar problem for FSP.71. I was comparing this with results shared by Tom for AT&T and putting the uncertainty corodinates from him I got the NLCD code exact as 52. But all surrowinding pixels are 42 and in AT&T computation it is 42 making a huge change in clutter loss calculation.
It is not specified anywhere but I am planning on taking the surrounding pixels and voting for the majority so that one off cases like this do not cause problems especially for use cases like this when uncertainty point is near the edge.
https://docs.google.com/presentation/d/1f655lJkrhLSxkTe5t4MBdH77AWJtHXnj8q4FOZtq0so/edit?usp=sharing
- @sergebdt even the above handling is different. The 2019 NLCD Code downloaded from the US Govt site is different from the one provided in the WINNF repository. I have raised a new issue regarding this as it is affecting many test vectors NLCD Code in WiNNF repo is different from US Govt NLCD 2019 data #64
As I said previously, the SAS project normalized on the 2011 NLCD (released in 2014). If the AFC project is using another reference, then the source data will obviously be different -- and any results you expect will be different than one obtained with SAS NLCD database. You should make sure that the AFC project has correctly and clearly normalized on one set of data, and use that. Has the AFC project clearly defined in its specification which data to use for the compliance tests ?
No, in the WINF-TS-1014 v4 it is mentioned to use https://www.mrlc.gov/data, but not which date.It should mean then to use the latest always for production I guess. But the question is now only with regard to matching the test vector output where as you mentioned this should be clearly mentioned
[n.18] National Land Cover Database (NLCD) data for CONUS, Alaska, Hawaii and Puerto Rico, available at https://www.mrlc.gov/data
In https://github.com/Wireless-Innovation-Forum/Common-Data/ it is mentioned that it holds the Common Data for the SAS and AFC project as well.
@AEgbert could you please clarify if the test vector results were calculated from NED form the WINNF repository or is it from US Govt 2020 land cover ?
While checking the discrepancy between calculated and expected values of FSP.23 came across a point (among so many uncertainty points) where on debugging found a difference between the WINNF provided data and the latest US Govt data ( NLCD_2021_Land_Cover data)
This is similar to issue #62
If I use the NLCD tiles from WINNF https://github.com/Wireless-Innovation-Forum/Common-Data/tree/master/data/nlcd nlcd_n34w119_ref.zip
I get NLCD code =22 for the uncertainty point and PSD does not match the test vector.
When I downloaded the Tiles from the US Govt https://www.mrlc.gov/viewer/ for 2021 land cover NLCD_x3tyHGqm37Jn4p5rSP6_LA7.zip and loaded this in QGIS and checked via the value tool it shows the difference between the WINNF data as can be seen below
More details
Using the downloaded 2021 land cover data
Name | NLCD_2021_Land_Cover_L48_20230630_x3tyHGqm37Jn4p5rSP67 Path | /ssd/qgis/NLCD_x3tyHGqm37Jn4p5rSP6_LA7/NLCD_2021_Land_Cover_L48_20230630_x3tyHGqm37Jn4p5rSP67.tiff
This causes FSP.23 test case to fail for
6640.0-6650.0
frequency rangeNLCD Code selected is Suburban among the uncertainty points as it has the lowest loss; and since loss is less, PSD also less
However, if we set this as URBAN we get the proper values