Question about SampleSelection
I have a question about sample selection : it seems there is not a good balance between the two Sentinel2 tiles we are using. Please note that we are not using a multi-region model : we use one classifier for both tiles and the whole island.
On this screenshot, we clearly see that samples for ocean (class 15) are only taken from the south tile (20PPC). We observe the same behaviour for the forest (class 10) : samples mostly (90%) come from the south of "Basse-Terre" (South-West part of Guadeloupe) whereas there are a lot of forest in the other regions. For the other classes, the balance is quite good between the two tiles.
Here are the sample selection parameters we used :
argTrain: { classifier :'rf'
options :' -classifier.rf.min 5 -classifier.rf.max 25 '
# option to configure sampling selection
sampleSelection : {"sampler":"random",
"strategy":"byclass",
"strategy.byclass.in":"/home/qt/tanguyy/eolab/tanguyy/OSO_Guadeloupe/IOTA2/SampleSelection/nb_pixels_per_class.csv"
}
}
Do you have an idea of how to improve the balance in our sample selection ?
It seems that for these ocean & forest classes, it started by the south tile (the 1st listed in Iota2 conf) and stopped after having reach the number of samples wanted.
Maybe a "periodic" strategy should be most suited for this purpose ?