Science Service System

Summary of Proposal MTH3093

TitleEnabling Widespread Adoption of Commercial SAR with Automated InSAR Signal Detection Algorithms
Investigator Woods, William - Stanford University, Electrical Engineering
Team MemberNo team members defined
SummaryWe generally expect SAR scientists to manually searchinterferograms for deformation signals. Reliance on human labor for signalidentification prohibits widespread adoption of commercial SAR. We will automate InSAR signal identification in order toencourage the widespread adoption commercial SAR. We will trainLocally-Weighted Regression learning algorithms to automatically identifydeformation signals in fully formed inteferograms and correlation maps. Funding SCRAP is currently un-funded. A common interest in remotesensing, entrepreneurship, philanthropy, and space research unites the team. Phase I Introduction Phase I identifies potential applications forhigh-resolution ultra-fast revisit SAR. Here we present a brief summary of fivetarget applications. Applications Disaster Management Application: Ultra-fast revisit times allow SAR users tomonitor progress of natural and man-made disasters in near real-time. Maritime Surveillance Application: Defense and law enforcement agencies wish tomonitor global waterway traffic in real-time. Weather Monitoring Application: Commissioners, designers, builders, andmanagers of sustainable power plants wish to know local wind velocity andcloud-cover conditions to both select appropriate site locations as well asquantify operational productivity of these power plants. Natural Resource Management Application: Monitor variations in forest vigor, identifyregions damaged by weather or disease, delineate boundaries between forests andprotected habitats, and identify potential illegal activity on unprecedentedtime scales. Infrastructure Monitoring Application: Recent drought conditions underscore theimportance of California’s water infrastructure. Government officials want tomonitor health of pipes, canals, and aqueducts responsible for supplying waterto the state. Phase II Introduction We select a learning algorithm to capable of identifyingdeformation signals typically found in InSAR. We then select SAR scenes togenerate InSAR images on which to train and test the algorithm. Algorithm We select the Locally-Weighted Regression learningalgorithm for this purpose. The LWR learning algorithm ingests training InSARimages for which deformation signals have been identified by hand. Anidentified signal consists of a contiguous set of pixels and their valuesselected from the training image. Once the neural network has learned thecharacter of those signals it can then automatically identify similar signalsin new images without human intervention. The probability of true-positive,false-positive, and true-negative quantify our algorithm’s effectiveness. Data Requirements We request minimum 32 focused images acquired in StripmapMode. We request multi-temporal images from a single coastal urban scene, suchas the San Francisco bay area, that are captured within a maximum time range ofone year and are compatible for interferometric processing. Start and stopdates may fall anywhere within the elapsed lifetime of TerraSAR-X. Phase III Introduction In Phase III we train our algorithms with TerraSAR-Ximages specified in Phase II. We calculate probability of success and evaluatealgorithm performance. We format and present our results to project advisorsand sponsors. Required Resources We use Professor Howard Zebker’s InSAR processor, whichwas built by the Radar Remote Sensing Group at Stanford University, to generateall possible pairs of InSAR images for the scene requested in Phase II. Wesupplement RRSG’s InSAR processor with a custom front end that ingestsTerraSAR-X images and formats them for processing. The team has priorexperience processing TerraSAR-X images. We identify signals in training images by hand. Personal computing equipment with Matlab is sufficientfor training LWR learning algorithms and storing SAR image products. Noadditional resources required.

Back to list of proposals

© DLR 2004-2016