140 likes | 280 Views
Computer-based identification and tracking of Antarctic icebergs in SAR images Tiago A.M. Silva, Grant R. Bigg Department of Geography, University of Sheffield, 2004. A Review By Aliah Chowdhury. February 2006. Introduction. Purpose Antarctic Icebergs influence climate and ocean changes
E N D
Computer-based identification and tracking of Antarctic icebergs in SAR imagesTiago A.M. Silva, Grant R. BiggDepartment of Geography, University of Sheffield, 2004 A Review By Aliah Chowdhury February 2006
Introduction • Purpose • Antarctic Icebergs influence climate and ocean changes • Application • Tracking iceberg activity by some efficient mechanism • Methodology • Use previous image processing techniques on SAR images to identify icebergs • Extend method to include tracking of temporal changes
Outline • Previous Approaches • The Extension in this Study • The Algorithm • Implementation & Results • Discussions by the Author • Evaluation • Final Words & Questions
Manual Ship-borne observations Limited Information No ability to track Unable to find all icebergs Satellite Imagery Difficult to identify Inefficient visual tracking Unable to measure changes Computerised Initial use of IP Under sampling Removes smaller objects Merges close objects Improved IP Technique Edge detection Watershed Segmentation Separates objects Extracts shapes Previous Approaches toIceberg Observation • All of these have their own merit and serve different applications
This Study: Extend Image Processing Approach to Track Iceberg Activity
The Extended Feature • Requirements • Detect and identify smaller icebergs • Separate from coastal ice sheets • Isolate icebergs from other objects • Track temporal changes & movement • Most Suitable Method • Two Stage Classification • Simple ‘Parametric’ Classifier • Unsupervised and Quantitative
The Algorithm (1) • Pre-processing • Block average with 2x2 window, under sample by 2 • Increases pixel width 12.5m to 25m: within ~26m res. • Coastal Masking • Reduce res. by 4, block average with 4x4, under sample • Operator marks polygons within each coastal region • Watershed segmentation; interpolation; binary masking • Segmentation • Multiresolution Ratio of Averages Filter; varied window sizes • Ratios in 4 direction windows normalised to build edge map • Watershed segmentation with threshold of -10dB • Oversegmentation solved by merging rule with minimum: 15% shared contour, |2dB| intensity difference
The Algorithm (2) • Classification – Iceberg Classifying The identification of icebergs is achieved using a simple parametric classifier with the following limits: • Classification – Iceberg Tracking Matching icebergs in images at different times/locations. • Paired & ranked by size similarities, then shape resemblance. • Size matching by minimum distance classifier between feature vectors composed of √area and major/minor axes. • Objects with <500m distance tested for shape matching • 1D distance/direction vector resampled every 5° into shape vector • All similar objects compared by shifting vector, then classified
ESA PRI/IMP Image Image Extraction and Calibration Offline Processing Edge Mapping Coastal Masking Yes No Polygon Markers Mask Coast? Watershed Segmentation Hierarchical Watershed Select & Apply Threshold Watershed Segmentation Online Processing 1st Selection & Merging Segmentation Iceberg Classification Review Classification KEY Area Estimate Correction In/Out User Intervention Iceberg Map & Database
Conditions & Results Conditions • 3 wintertime intensity images • Local methods of segmentation • 3x3 & 5x5 window sizes • Only σ° > -8dB classified icebergs • Manual classification of icebergs • Brighter than background • Bright rim closer to sensor • Shadow away from sensor • Angular corners • Manual tracking v computerised Results • Over 90% of iceberg correctly segmented over all images • ~70% of pixels classified correctly • Between 60-100% of all matches detected accurately • Misses or mismatches due to incorrect segmentation • Better detecting smaller icebergs • But misses usually smaller icebergs • 10-13% area underestimated • Implementation • MATLAB Image Processing Toolbox, images in FFT • ESA BEST Software for object analysis
Discussion by Authors • σ° values for background, sea ice and icebergs overlap in images; • Reduced values for icebergs can be due to recent phenomena; • Higher values for background due to sea ice concentration. • Looking angle & pass direction differ between acquisitions. • Segmentation process affected by slopes of icebergs at edges. • Robust rotation feature untested • Applicable to many areas of study • Extendable to other Imagery
Evaluation of Study • Use of previous methods constructive and good for validation • Classification stage most successful; failures mostly due to unavoidable circumstances • Utilises features and IP techniques to great effect • Blend of qualitative and quantitative measurements • The study admits inconsistency and needs to address more rigorous validation methods • Useful potential, widely extendable, though needs refinement
Final Words & Questions THANK YOU!