Change detection by deep neural networks for synthetic aperture radar images
Liao Frank, Koshelev Elizabeth, Milton Malcolm, Yuanwei Jin, Enyue Lu · 2017 International Conference on Computing, Networking and Communications (ICNC) · 2017
In this Research Experience for Undergraduate (REU) project, we develop and implement deep neural network algorithms for change detection of synthetic aperture radar (SAR) images. Deep neural networks represent a powerful data processing methodology that integrates recent deep learning techniques on neural network computing frameworks to undercover underlying features and structures of observational data. The classic change detection method for SAR images is through the difference image analysis method, i.e., filtering the noise in each before-change and after-change image and then identifying the changes between the two images. Although well researched, the difference image method requires significant pre-processing and has difficulty with applications that require high accuracy and flexibility. The proposed deep neural networks create a change detection map from original SAR images directly without generating difference images, thus providing a novel framework for change detection of complicated SAR images where speckle noise is also present. We conduct numerous experiments on artificial images with added speckle noise and real-world synthetic aperture radar images.