- A model trained using patches from the original image size without resizing it down to a width of 2048 and generated tiles using an overlap of 80%