Giters

chathumal93 / Pytorch-UNet-Flood-Segmentation

This repository contains a PyTorch implementation of a U-Net model for segmenting water areas (flood and permananet water) in Sentinel-1 satellite images.

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

too few test samples

makovez opened this issue · comments

commented

Hi, I think your result of 0.87 IoU might be influenced by the few samples number in test set:
Your overall split train/test is 97.6% train and 2.4% test.
Have you tried decreasing train and increasing test samples ?

commented

Train, validation and test percentages are respectively 88%, 10% and 2%. Test dataset is small as you have observed and the test dataset is selected in a way that the water percentage ranges from 0-100 percent. I haven't tried reducing the training dataset, since the whole dataset is small.