Training data dimensions in SNAP


I am using a high resolution image to collect reference data, which I will then use to classify my Sentinel 1 and 2 imagery. I realized that the training data have sizes like 3m in length and maybe 1m or so in width. Could this present a problem when classifying the sentinels data since they are at least 10m resolution?

Hi @pamochungo,

In that case, your training data (I assume you are creating polygons) will overlap with one single pixel in Sentinel-1 and Sentinel-2. It is not a problem a main issue (depending on the classes you want to classify) but you have to make sure you have enough training data (pixels).

What if you create bigger training data? What do you want to classify?


Thanks M,

Apart from classifying the normal broad land cover classes, I need to classify Hedges, which are very narrow and cannot be more than 3m wide.



as long as your areas contain spectral features, they can serve as training input. But keep in mind that some classifiers need larger sample sizes. Random Forest, for example, needs to split the training data into random subsets, in order to work thoroughly. If your class is only represented by one pixel, this randomization won’t have the proper effect.

If you area searching for fine structures maybe the spectral unmixing is the better choice?

I have never used spectral unmixing. Do you have a link?



Topics in this forum:

Please also see the SNAP internal Help section on spectral unmixing.