Waterbody mapping using image segmentation
This project was class project done by me on my computer vision class in Brown University.
Water is an essential part of a living ecosystem, as it is necessary for the survival of all living beings. As a result, many scientific tasks require up-to-date information on spa- tial distribution of water; to be able to track trends in water resources over time. To do this, it is common to look at wa- ter maps, which are masks of satellite images where each pixel is either water or non-water: In this project, we wanted to automatically generate water maps using deep learning techniques. We collected Landsat 8 imagery with existing water maps to train a U-Net model, which has previously been shown to achieve high accuracy in image segmentation tasks. After training for 100 epochs, our model obtained accuracy of over 99% on the test set. While it was able to identify large regions of water; some of the small water bod- ies and finer details were often misclassified by our model. Counterintuitively, data augmentation did not produce better results and instead reduced the accuracy. However, these results are still very promising and open the door to quick and easy water mapping in the future. Go to this github link to see source code
