Hello, I noticed that you mentioned in "Participate>>Get Data>>Land Cover Labels" that Savanna (class #3) and Snow/Ice (class #8) are not included in the validation and test sets. I have two questions:
1. I computed the IGBP values of all the pixels in the validation set. It turns out that there exist some pixels which belong to Savanna (class #3). Is there something wrong?
2. Are we supposed to ignore Savanna (class #3) and Snow/Ice (class #8), and only predict the other classes in the validation set?
>1. I computed the IGBP values of all the pixels in the validation set. It turns out that there exist some pixels which belong to Savanna (class #3). Is there something wrong?
Yes, there are some pixels belong to Savanna but Savanna is not included in high-resolution ground truth of both validation and test sets.
>2. Are we supposed to ignore Savanna (class #3) and Snow/Ice (class #8), and only predict the other classes in the validation set?
Exactly.
I got it. Thank you!
Posted by: WillChan @ Feb. 3, 2020, 5:58 a.m.It's important to understand that for the validation set you get the same data as for the SEN12MS dataset:
- Sentinel-1 (two channels)
- Sentinel-2 (13 bands)
- low-resolution (resolution: 500m), MODIS-derived land cover (4 maps, one of which can be re-mapped to the simplified IGBP scheme using the code provided in the "data loader")
The evaluation, however, will be carried out against high-resolution (resolution: 10m) land cover maps, which have been created only for this contest.
While the 500m input land cover does have savanna class on board, the 10m high-resolution data doesn't. Therefore, one has to take specific care about what to do with the Savanna class, especially since this class does not make a lot of sense in (a) global (b) high-resolution data in the first place.
While there is also no Snow/Ice in the high-resolution reference data, there is no need to bother with this class, as there is also only very few examples in the original SEN12MS data. At least we have not faced any problems with this class.
Posted by: MSCHMITT @ Feb. 3, 2020, 7:27 a.m.Got it. Thanks!
Posted by: WillChan @ Feb. 3, 2020, 7:38 a.m.