Water Body Segmentation

Introduction

In this project, I train and evaluate a U-Net (implemented in Keras) to segment water bodies from satellite images. A baseline U-Net acheived a dice-score of 0.85. I have yet to add an augmentation step.

Data

Data was downloaded from https://www.kaggle.com/franciscoescobar/satellite-images-of-water-bodies. It consisits of two folder (Images and Masks), each containing 2841 JPEG images (with corresponding names). They are of various image shapes. This means I needed a resizing step in my image generator.

Training and Required Packages

I used Google Colab (and a GPU runtime) to train on 1988 images, which took about 2 seconds/batch of 3 images. Required libraries

• numpy
• os
• keras
• cv2
• matplotlib

Files

dice.py

• dice coefficient (metric) and dice loss (custom loss)

generator.py

• image generator which reads images and masks from Data/Images and Data/Masks
• resizes to desired shape (I used 256,256)
• yields batches upon reaching desired batch size

models.py

• here I define a U-Net and a smaller segmentation model (U-Net skinny)

train.py

• creates.compiles model from models.py
• train/validaton/test split (70/20/10) assuming all one folder each for images and masks
• train with callbacks (EarylStopper, CSVLogger, and checkpointer)

predict.py

• display image, true mask and predicted mask
• OR
• display image and predicted mask if there is no correspoinding ground truth mask

evaluate.py

• loads desired model from results folder
• evaluates on validation data

Results

Baseline model (unet from models.py) acheived a dice score of 0.85. Below are some examples of the images (from the validation set), the true mask and the predicted mask.

	Image	Mask	Prediction
Example 1
Example 2	Content Cell