5A_ 2_Developing a statistical methodology to improve classification and mapping of seabed type from deep water multi-beam echo sounder data

Developing a statistical methodology to improve classification and mapping of seabed type from deep water Multi-Beam Echo Sounder (MBES) data Helen Caughey, Kazi Ishtiak Ahmed, Paul Harris, Peter Hung, Urška Demšar, Sean McLoone, A Stewart Fotheringham, Xavier Monteys, Ronan O’Toole Presented by: Helen Caughey National Centre for Geocomputation National University of Ireland, Maynooth helen.m.caughey @ nuim.ie GISRUK 2010, University College London, 15th April 2010

Overview of Work Collaboration between StratAG and G eological S urvey of I reland/ M arine I nstitute/INFOMAR programme. Focuses on seabed type mapping from data acquired with a M ulti- B eam E cho S ounder (MBES). The work consists of 2 main phases; PHASE 1 ; the use of spatial and statistical analysis to decide which datasets can be integrated and how. PHASE 2 ; consists of finding the best automatic classification methodology for integrated data and evaluation of classification quality against ground truth and/or existing seabed maps.

MBES Working Principles MBES raw data – backscatter measurements Image compensation to remove effects of sonar angle and range 132 Full Feature Vectors (FFVs, statistical descriptors) for each patch Interpolated acoustic image (pixels) Interpolated acoustic image (patches) Data acquired on receiver Narrow beam width Wider swath Both Bathymetry and backscatter Port Starboard

Data used was collected over 3 years (2000, 2001 & 2002) and at 3 different pulse lengths (2000, 5000 & 15000 ms). Systems recalibrated at the beginning of each survey season. Each survey yielded a number of backscatter features, five of which we have extracted; Q (a quantile measure), P (the ‘pace’ textural feature), C (a ‘contrast’ feature), M (the mean) and S (the standard deviation). Bathymetry data is also available for the same zone 3 area.

Phase 1: Data integration based on spatial and statistical analysis; Is data integration possible? Removal of pulse length 2000 from analysis

Phase 1: Data integration based on spatial and statistical analysis; Is data integration possible?

Phase 1: Data integration based on spatial and statistical analysis; Is data integration possible? Identification of spatially overlapping data subsets Removal of pulse length 2000 from analysis 6 overlapping, spatially diverse, areas identified from PL 5000 & 15000 (each overlap also has multiple sub areas)

Phase 1: Data integration based on spatial and statistical analysis; Is data integration possible? Identification of spatially overlapping data subsets Removal of pulse length 2000 from analysis 6 overlapping, spatially diverse, areas identified from PL 5000 & 15000 (each overlap also has multiple sub areas) “ Pseudo Pairwise” overlapping data subsets were then identified and refined automatically in the R statistical computing environment

Phase 1: Data integration based on spatial and statistical analysis; Is data integration possible? Summary Statistics and distributions

Phase 1: Data integration based on spatial and statistical analysis; Is data integration possible? Summary Statistics and distributions Classical hypothesis testing (F, t and K-S tests) Differences in Spatial Autocorrelation; investigated using estimated variograms ( M ethods o f M oments Variograms) and modelled variograms ( Re stricted M aximum L ikelihood Variograms) Correlation and error analyses; Scatterplots & weighted error diagnostics, Weighted correlations, Robust & weighted correlations

Phase 1: Data integration based on spatial and statistical analysis; Is data integration possible? Unlikely these datasets should be joined; OL4 OL5 OL7 OL8 Likely that these datasets could be joined; OL6 OL9

Phase 1: Data integration based on spatial and statistical analysis; Is data integration possible? Focused scale investigations

Outcomes & Recommendations (Phase 1) OUTCOMES FOR PHASE 2 Based on these statistical tests and results we would suggest that the datasets should not be joined together. This does not however rule out joining the dataset if they are rescaled based on the knowledge we have gained about the datasets and their relationships. If they are rescaled then we can incorporate a measure of the uncertainty into the final classification. RECOMMENDATIONS TO GSI/MI/INFOMAR In the past they have joined datasets assuming that there is no variance between them. These tests have shown this is not the case and therefore we have recommended that datasets not be joined together without a method of validation prior to joining. More care needs to be taken in the recalibration of the survey systems each year. More emphasis needs to be placed on some level of overlap from one year to next in order to have the necessary data to validate any joins.

Phase 2: Data classification processes Proceed with classification attempts without joining the data. PCA and K means as preliminary approach. Other neural networks and quality threshold clustering. Next step – Data Classification

Thank you! Questions? Acknowledgements: Contacts: [email_address] ncg.nuim.ie www.stratag.ie

5A_ 2_Developing a statistical methodology to improve classification and mapping of seabed type from deep water multi-beam echo sounder data

More Related Content

What's hot (19)

Viewers also liked (9)

Similar to 5A_ 2_Developing a statistical methodology to improve classification and mapping of seabed type from deep water multi-beam echo sounder data (20)

More from GISRUK conference (20)

5A_ 2_Developing a statistical methodology to improve classification and mapping of seabed type from deep water multi-beam echo sounder data