I have dataset of multispectral images of land areas where the most of images are less than 5 by 5 pixels in size. I read that usual way to deal with small images is to upscale it to fixed size lets say 32x32 pixels and then feed the cnn by them. Because of multispectral images I would like to avoid upscaling images because it will bring too much noise to data where values of intensities of pixels are very important and it will destroy ground truth.Firstly I would like to know if it is possible to use such small images for cnn ? Also I am wondering if padding such small images with zeros is good approach. I am planning to add my ground truth as extra channel for cnn to know that zero padding is not important at all.
Or is there any approach to deal with very small images for cnn ?
Thanks in advance for your suggestions.
EDIT
I am working with multispectral Sentinel 2 images in 20m resolution meant as 3x3 pixel image is 60x60 m in extent in real. Most relevant information is in values of intensities of nine different bands.
For example here is 3x3 image of cemetery in blue and vegetation red edge band.
