Automatic Generation of Water Masks of the Mekong Delta, Vietnam Using a Fully Convolutional Network Trained Using Synthetic Aperture Radar Images.

Abstract:

As student interns at ASF, the purpose of this project was to identify water in the Mekong Delta, Vietnam. This project is an extension of the Alaska Satellite Facility’s (ASF) Fully Convolutional Network (FCN), created by student interns in 2019, to identify water using Sentinel-1 Synthetic Aperture Radar (SAR). Our goal is to improve the techniques and accuracy of water masking for this area using SAR and the FCN developed by ASF.

First, Sentinel-1 data was radiometrically terrain corrected (RTC) using ASF’s Hybrid Pluggable Processing Pipeline (HyP3). Next, training datasets were created using a statistical algorithm. Training datasets were refined by comparing the statistical results and those from the FCN created by ASF in 2019, manually choosing the best fitting water mask. QGIS was used to assist in identifying areas of water. Over 30 GB of datasets were created. All operations were performed using Amazon Web Services (AWS).

It was challenging to create training datasets since the landscape of the Mekong Delta changes drastically throughout the year due to rainfall and flooding. The area is intentionally flooded for agricultural purposes, making it difficult to accurately differentiate between areas of water versus land, such as rice fields. It was found that using data during the dry season helped distinguish these areas more clearly when making datasets.

The preliminary training ran for 100 epochs on a groomed dataset of Mekong Delta granules, with dates ranging from August 2019 to May 2020. The resulting FCN showed an accuracy of 99% upon testing. SAR images that were input through the water mask generation process visually resulted in a mask representative of the water bodies on the Mekong Delta. It was also found that the rice paddies could easily be recognized by the FCN. Overall, these results are encouraging, as the FCN was generally able to distinguish water from land.