In Patch wise training, the CNN is trained on a small patch of the image instead of the whole image. For example, if the input image is of size 6000-by-6000 pixels, it would be difficult to train a CNN on it. You could downscale the image but then you would lose relevant information. Instead, if you perform patch based semantic segmentation, the model would pick up patches from the image, say of size 200-by-200 pixels and train on those patches as if they were separate images.
This feature retains information as well as provides good accuracy.A link to a paper discussing the relevance of patch-based CNN can be found here.