Using FCNN for multi-class semantic segmentation trained on single class labeled image data

Question

I am working on project where main task is semantic segmentation of land cover and another objects in Sentinel 2 multi-spectral images. Currently I posses datas