Sigmoid or Softmax for multichannel conv2d Output in image segmentation

Question

I am training my U-architecture for pixelwise classification. I have 9 classes and all the classes are mutually exclusive. Each label is represented with differ