Water Regions Extraction From Radar Imagery Using a Neural Network
Abstract
An artificial neural network concept is explored and developed for detecting and extracting water regions from radar imagery. A backpropagation neural network consisting of three layers of processing elements (PEs) is selected for this application. The input layer is composed of nine PEs that are arranged to process a single pixel at the same time. Two PEs in a hidden layer are sufficient for delineating water from other terrain categories. A single output PE with a thresholding function classify effectively all test images into two terrain categories: water and non-water. Large-scale Synthetic Aperture Radar (SAR) images, containing 512 by 512 pixels, were used as the test images for this experiment. Two blocks of water regions totalling 2,048 pixels were extracted for training the network. All the water pixels were classified correctly, while more than 99 percent of the non-water pixels were correctly classified.
Document Details
- Document Type
- Technical Report
- Publication Date
- Dec 16, 1992
- Accession Number
- ADA259517
Entities
People
- Pi-fuay Chen
- Tho Con Tran
Organizations
- Army Geospatial Center