330 likes | 481 Views
A Novel Approach to the Automatic Detection of Subsurface Features in Planetary Radar Sounder Signals. Adamo Ferro Lorenzo Bruzzone. E-mail: adamo.ferro@disi.unitn.it Web page: http:// rslab.disi.unitn.it. Outline. Introduction. 1. Aim of the Work. 2.
E N D
A Novel Approach to the Automatic Detection of Subsurface Features in Planetary Radar Sounder Signals Adamo Ferro Lorenzo Bruzzone E-mail: adamo.ferro@disi.unitn.it Web page: http://rslab.disi.unitn.it
Outline Introduction 1 Aim of the Work 2 Statistical Analysis of Radar Sounder Signals 3 4 Automatic Detection of Basal Returns Conclusions and Future Work 5 A. Ferro, L. Bruzzone
Introduction • Planetary radar sounders can probe the subsurface of the target body from orbit. • Main instruments: • Moon: ALSE and LRS • Mars: MARSIS and SHARAD • Their effectiveness lead to the proposal of new orbiting radar sounders, also for Earth science: • IPR and SSR for the Jovian Moons[1] • GLACIES proposal for the Earth[2] • Radar sounder data have been analyzed mostly by means of manual investigations. v Platform height Nadir Across track Range (depth) [1] L. Bruzzone, G. Alberti, C. Catallo, A. Ferro, W. Kofman, and R. Orosei, “Sub-surface radar sounding of the Jovian moon Ganymede,” Proceedings of the IEEE, 2011. [2] L. Bruzzone et al., “GLACiers and Icy Environments Sounding ,” response to ESA’s EE-8 call, 2010. Example of radargram (SHARAD) A. Ferro, L. Bruzzone
State of the Art • Past works related to the automatic analysis of radar sounder data regard the analysis of ground-based or airborne GPR signals. • Different frequency ranges. • Better spatial resolution. • Detection of buried objects (e.g., mines, pipes) which show specific signatures (e.g., hyperbolas). • Investigation of local targets vs. regional and global mapping. • Planetary radar sounding missions are providing a very large amount of data. • In order to effectively extract information from such data automatic techniques can greatly support scientists’ work. A. Ferro, L. Bruzzone
Ground processing Preprocessing Information extraction ProposedProcessing Framework Level 2 products Level 3 products Map of interesting areas Labels Raw data Level 1products Icy layers position 3D tomography of icy layers Basal returns position Ice thickness map ... ... ... Other inputs(e.g., ancillary data, clutter simulations) A. Ferro, L. Bruzzone
Aim of the Work • Development of a processing framework for the automatic analysis of radar sounder data. • Statistical analysis of radar sounder signals. • Characterization of subsurface features. • Basis for the development of automatic techniques for the detection of subsurface features. • Automatic information extraction from radargrams. • First return. • Basal returns. • Subsurface layering. • Discrimination of surface clutter. A. Ferro, L. Bruzzone
Aim of the Work • Development of a processing framework for the automatic analysis of radar sounder data. • Statistical analysis of radar sounder signals. • Characterization of subsurface features. • Basis for the development of automatic techniques for the detection of subsurface features. • Automatic information extraction from radargrams. • First return. • Basal returns. • Subsurface layering. • Discrimination of surface clutter. A. Ferro, L. Bruzzone
Dataset Description • SHARAD radargrams • Number of radargrams: 7 • Area of interest: North Polar Layered Deposits (NPLD) of Mars • Resolution: 300 × 3000 × 15 m (along-track × across-track × range) -2500 m SHARAD radargram 1319502 -5500 m A. Ferro, L. Bruzzone
Proposed Approach: Statistical Analysis • Goal: • Understand the statistical properties of the amplitude distribution underlying the scattering from different target classes. • Definition of targets: • NT: no target • SL: strong layers • WL: weak layers • LR: low returns • BR: basal returns SHARAD radargram 1319502 A. Ferro, L. Bruzzone
Proposed Approach: Statistical Analysis • Tested statistical distributions (amplitude domain): • Rayleigh: simplest model, scattering from a large set of scatterers with the same size. • Nakagami: amplitude version of the Gamma distribution, has the Rayleigh has a particular case. • K: models the scattering from scatterers not homogeneously distributed in space, which number is a negative binomial random variable. • Distribution fitting performed via a Maximum Likelihood approach. • Goodness of fit tested by calculating the RMSE and the Kullback-Leibler distance (KL) between the target histogram and the fitted distribution. Mean power Amplitude Shapeparameter Shapeparameter A. Ferro, L. Bruzzone
Proposed Approach: Statistical Analysis, Fitting SHARAD radargram 1319502 No target Strong layers Weak layers Low returns Basal returns Summary A. Ferro, L. Bruzzone
Results: Statistical Analysis • Best fitting distribution: K distribution • The parameters of the distribution describe statistically the characteristics of the target. • Noise can be modeled with a simple Rayleigh distribution. A. Ferro, L. Bruzzone
BR seed selection • for m=2 to M • Region growing • Thresholding • Thresholding • BR seed selection • Estimation of BR statistics • Calculation of KLHN • Region selection • First return detection • BR map generation • Region growing Proposed Approach: Automatic Detection of BR Inputradargram • KL1 • Initial BR map • KLHN map • KLm BR map A. Ferro, L. Bruzzone
for m=2 to M • Region growing • BR seed selection • Thresholding • Thresholding • BR seed selection • Estimation of BR statistics • Calculation of KLHN • Region selection • First return detection • BR map generation • Region growing Proposed Approach: Automatic Detection of BR • Frame-based detection of the first return. • Map of the KLHN: • Calculated for the subsurface area using a sliding window approach. • It represents a meta-level between the amplitude data and the final product. Inputradargram Local histogram Estimatednoisedistribution • KL1 • Initial BR map • KLHN map • KLm SHARAD radargram 1319502 BR map A. Ferro, L. Bruzzone
for m=2 to M • Region growing • BR seed selection • Thresholding • Thresholding • BR seed selection • Estimation of BR statistics • Calculation of KLHN • Region selection • First return detection • BR map generation • Region growing Proposed Approach: Automatic Detection of BR • Frame-based detection of the first return. • Map of the KLHN: • Calculated for the subsurface area using a sliding window approach. • It represents a meta-level between the amplitude data and the final product. Inputradargram Local histogram Estimatednoisedistribution • KL1 • Initial BR map • KLHN map • KLm SHARAD radargram 1319502 BR map A. Ferro, L. Bruzzone
Estimation of BR statistics • BR map generation • First return detection • Region selection • Calculation of KLHN • Region growing • BR seed selection • Thresholding • Thresholding • BR seed selection • Region growing • for m=2 to M Proposed Approach: Automatic Detection of BR • Frame-based detection of the first return. • Map of the KLHN: • Calculated for the subsurface area using a sliding window approach. • It represents a meta-level between the amplitude data and the final product. Inputradargram Local histogram Estimatednoisedistribution • KL1 • Initial BR map • KLHN map • KLm SHARAD radargram 1319502 KLHN map BR map A. Ferro, L. Bruzzone
for m=2 to M • BR map generation • Region selection • Region growing • BR seed selection • Thresholding • Estimation of BR statistics • Region growing • BR seed selection • Calculation of KLHN • First return detection • Thresholding Proposed Approach: Automatic Detection of BR • Selection of the regions with the highest probability to be related to the basal scattering area. • The initial BR map is created using a region growing approach based on level sets which starts from the seeds and moves on the KLHN map. Inputradargram Propagation Curvature Level set function • KL1 • Initial BR map • KLHN map • KLm KLHN map Initial BR map BR map A. Ferro, L. Bruzzone
Region growing for m=2 to M BR seed selection Thresholding Thresholding BR seed selection Estimation of BR statistics Calculation of KLHN Region selection First return detection BR map generation Region growing Proposed Approach: Automatic Detection of BR • The initial BR map is used to estimate the statistical distribution of the amplitude of the BR samples. • The procedure is repeated iteratively using lower threshold ranges for the KLHN map. • The new regions created during the iterations which are not statistically similar to the estimated BR distribution are deleted. Inputradargram KL1 Initial BR map KLHN map KLm Initial BR map BR map A. Ferro, L. Bruzzone
Region growing for m=2 to M BR seed selection Thresholding Thresholding BR seed selection Estimation of BR statistics Calculation of KLHN Region selection First return detection BR map generation Region growing Proposed Approach: Automatic Detection of BR • The initial BR map is used to estimate the statistical distribution of the amplitude of the BR samples. • The procedure is repeated iteratively using lower threshold ranges for the KLHN map. • The new regions created during the iterations which are not statistically similar to the estimated BR distribution are deleted. Inputradargram KL1 Initial BR map KLHN map KLm Step 2 BR map A. Ferro, L. Bruzzone
Region growing for m=2 to M BR seed selection Thresholding Thresholding BR seed selection Estimation of BR statistics Calculation of KLHN Region selection First return detection BR map generation Region growing Proposed Approach: Automatic Detection of BR • The initial BR map is used to estimate the statistical distribution of the amplitude of the BR samples. • The procedure is repeated iteratively using lower threshold ranges for the KLHN map. • The new regions created during the iterations which are not statistically similar to the estimated BR distribution are deleted. Inputradargram KL1 Initial BR map KLHN map KLm Step 3 BR map A. Ferro, L. Bruzzone
Results: Automatic Detection of BR SHARAD radargram 1319502 SHARAD radargram 0371502 SHARAD radargram 1292401 SHARAD radargram 1312901 A. Ferro, L. Bruzzone
Results: Automatic Detection of BR • The performance of the technique has been measured quantitatively. • Selection of 3000 reference samples randomly taken in areas of the radargram where BR returns are (or are not) visible. • Counted the number of samples correctly detected as BR (or not BR) returns. A. Ferro, L. Bruzzone
Results: LayerDensityEstimation SHARAD radargram 052052 Automatic detection of linear interfaces Interface density map A. Ferro, L. Bruzzone
Conclusions • Developing a processing framework for the analysis of radar sounder data. • Statistical analysis of radar sounder signals. • It can support the analysis of the radargrams. • Different statistics / different targets. • Generation of statisticalmapsusefulto drive detectionalgorithms. • Novel technique for the automatic detection of the basal returns from radar sounder data using statistical techniques. • Effectively tested on SHARAD radargrams. • Possible applications: estimation of ice thickness, detection of local buried basins or impact craters, 3D measurement of the scattered power, study seasonal variation of the signal loss through the ice. A. Ferro, L. Bruzzone
Future Work • Improvements of the proposed technique: • Estimation of local statistics using context-sensitive techniques for the adaptive determination of the local parcel size. • Develop a procedure for the automatic and adaptive definition of the parameters of the proposed technique. • Adapt the algorithm to airborne acquisitions on Earth’s Poles. • Other possible developments: • Integration of the automatic detection of linear interfaces and basal returns to higher level products. • Automatic detection and filtering of surface clutter returns from the radargrams. A. Ferro, L. Bruzzone
Thank you for your attention! • Contacts: • E-mail: adamo.ferro@disi.unitn.it • Website: http://rslab.disi.unitn.it A. Ferro, L. Bruzzone
BACKUPSLIDES A. Ferro, L. Bruzzone
AutomaticDetection of SurfaceClutter, Example SHARAD radargram 1386001 Coregistered surface clutter simulation Detected surface clutter map A. Ferro, L. Bruzzone
Automatic Detection of the NPLD BR, Results Example of application to a large number of tracks A. Ferro, L. Bruzzone -2300 20 -4000 0 Depth of detected BR fromdetected surface return [µs] Coverage of selected 45 tracks Mars North Pole topography [m] 180º 270º 90º 88º 86º 84º 82º 0º
Results: Automatic Detection of BR SHARAD radargram 1319502 SHARAD radargram 0371502 SHARAD radargram 1292401 SHARAD radargram 1312901 A. Ferro, L. Bruzzone
Modelparameters A. Ferro, L. Bruzzone