I want to perform supervised classification for river sand detection. I have 3 m spatial resolution Surface Reflectance multispectral imagery with B, G, R and NIR bands (Planet data). Which all features should I choose for classification input that would give best separation for my sample dataset?
In this case, your feature space is tiny - only four variables. Any reasonable classification method is designed to deal with a significantly larger feature space, so there is no need to reduce the data set.
As such, go ahead and use all the available data.
Many methodologies will also provide a 'ranking' of features, which will indicate how useful a given feature was for the classification. This information can then be used to guide you towards which spectral bands are the most important for your study.
I would recommend including all of the spectral bands (R,G,B, NIR) and a vegetation index such as normalized difference vegetation index (NDVI), which utilizes the red and near infrared bands. The NDVI index will especially help your classification algorithm separate vegetation and water classes from the river sand.