Development of Water Masks using Synthetic Aperture Radar via a Fully Convolutional Neural Network at the Alaska Satellite Facility
Abstract
Student interns at the Alaska Satellite Facility (ASF) have developed software, written in Python 3 and running on Amazon Web Services (AWS), to identify water in Sentinel-1 Synthetic Aperture Radar (SAR) satellite images using a Fully Convolutional Network (FCN). The output of the FCN is a bit mask indicating the location of water in a SAR image. This project is composed of two halves; the training data generator and the FCN.
The training data generator encompasses several steps. First, the ASF HyP3 web-based on demand SAR processing system performs radiometric terrain correction (RTC) on a collection of dual-pol SAR images. Next, a statistical algorithm is applied to generate an approximation of the "ground truth" bit mask. Next, each mask is labeled as containing water or not. Finally, the data is hand groomed by removing bad image masks and, where necessary, modified to increase accuracy. 40 SAR granules were processed resulting in 20 GB of high quality training data. This training data was uploaded to an AWS S3 bucket to provide convenient access. The FCN is built with Keras, an open-source neural network library running on top of TensorFlow. The FCN uses the Unet Architecture for semantic segmentation allowing each pixel to be labeled as either water or land. To allow this project to scale and meet ASF's needs, the FCN can be run on an AWS EC2 instance. The results are promising, with testing and training accuracy greater than 90%. However, validation and accuracy in machine learning can be deceptively difficult to interpret; in our case the "ground truth" is unknown. To address these questions, several metrics were collected and various data visualization techniques were used: charts of accuracy vs. epoch, loss vs. epoch, and Gifs of the masks developing. These and other aspects of the project will be discussed, as well as the potential for later versions of this software to be of interest to remote sensing scientists for land masking and speckle filtering.- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2019
- Bibcode:
- 2019AGUFMIN42A..06S
- Keywords:
-
- 1916 Data and information discovery;
- INFORMATICS;
- 1942 Machine learning;
- INFORMATICS;
- 1976 Software tools and services;
- INFORMATICS;
- 4329 Sustainable development;
- NATURAL HAZARDS