Principles of object-oriented image analysis Image mining and - PowerPoint PPT Presentation

Principles of object-oriented image analysis Image mining and knowledge-driven analysis in disaster risk management Dr. Norman Kerle

Lecture outline (1) Me @ (1) Me @ ITC ITC (2) Principles of (2) Principles of OOA & use OOA & use for DRM for DRM (3) Recent research (3) Recent research (4) Outlook & (4) Outlook & trends trends 2 Image Mining 2013 – Barcelonette - 28 August 2013

(1) Me @ (1) Me @ ITC ITC ITC/University Twente  ITC now faculty of Uni Twente  Houses the United Nations University- ITC Centre for Spatial Analysis and Disaster Risk Management  Training, education and curriculum development  Knowledge development and research collaboration  Advisory services  In collaboration with many partners www.unu-drm.nl 3 Image Mining 2013 – Barcelonette - 28 August 2013

Me @ ITC  Geographer with study background in Hamburg (D), Ohio State (US) and Cambridge (UK)  Since early 1990s work in the hazards & disaster field, with focus on remote sensing  PhD in volcano remote sensing (lahars)  Advanced image analysis, and focus on object-oriented analysis http://www.itc.nl/ooa-group http://www.itc.nl/about_itc/resumes/kerle.aspx 4 Image Mining 2013 – Barcelonette - 28 August 2013

(2) Principles of (2) Principles of OOA & use for DRM OOA & use for DRM Object-oriented analysis for disaster risk management DRM OOA Disaster risk  Different concepts  Expected losses (f[hazard, period])  Risk = Hazard * Vulnerability EaR * Amount (R=H*V)  EaR (elements at risk): not only physical  H: f(type, magnitude)  V: physical, social, economic, environmental, etc.  Amount: quantifiable?  Note: all elements of risk are spatial 5 Image Mining 2013 – Barcelonette - 28 August 2013

Basics of object-oriented analysis DRM OOA OOA  OOA is a form of image classification  Objects = segments (segmentation-based analysis, OBIA, GEOBIA) 1. step: segmentation: old concept (~1970s) – partition an image into  homogenous units  2. step: classification of those units 6 Image Mining 2013 – Barcelonette - 28 August 2013

Basics of object-oriented analysis DRM OOA OOA  Segmentation also at multiple scales, and using auxiliary information  Note: we do most OOA work in eCognition software Super-objects Classification level Sub-objects Pixel Any type of raster image, and thematic GIS layer 7 Image Mining 2013 – Barcelonette - 28 August 2013

Basics of object-oriented analysis DRM OOA OOA  Main difference over pixel-based methods: objects have extra features (spectral, geometric, contextual) = useful for classification  Allows use of feature and process knowledge 8 Image Mining 2013 – Barcelonette - 28 August 2013

Basics of object-oriented analysis  Pixel-based: landcover (spectal information); OOA - landuse  Challenge: we need detailed feature and process knowledge 9 Image Mining 2013 – Barcelonette - 28 August 2013

(3) Recent research (3) Recent research OOA for DRM  Our group addresses  Use of OOA for different hazards and risk elements  Different aspects of risk  Methodological work (better segmentation, feature and threshold selection) Domain focus Remote sensing for DRM Hazard Vulnerability EaR Risk Damage (Recovery) Technical Urban/ infra- Pictometry-/UAV- Refugee camps; Social Landslides/ structure based damage metrics for recovery focus erosion OOA (in eCognition) 10 Image Mining 2013 – Barcelonette - 28 August 2013

OOA for DRM  Focus on landslide work – find solutions to this type of mapping problem 11 Image Mining 2013 – Barcelonette - 28 August 2013

Problem: knowledge incorporation Hazard: Landslide work  Work with several PhD students and postdocs  work of Tapas Martha  conceptualization of a landslide  segmentation based on satellite data and elevation data  removal of false positives  classification of different landslide types OOA-based landslide mapping Martha et al., 2010 Full PhD thesis: www.itc.nl/library/papers_2011/phd/martha.pdf (Geomorphology) 12 Image Mining 2013 – Barcelonette - 28 August 2013

Problem: scale parameter Scale factor 20 Scale factor 50 Hazard: Landslide work  Problem: trial & error work  What segmentation parameters? One-fits-all?   Work on statistical optimization of segmentation  Balancing intra-segment homogeneity and inter-segment heterogeneity  Plateau objective function (POF) to select appropriate scale factors OOA-based landslide Objective segmentation mapping (POF) Martha et al., 2011 Martha et al., 2010 (IEEE TGRS) (Geomorphology) 13 Image Mining 2013 – Barcelonette - 28 August 2013

Problem: change detection Hazard: Landslide work  Ping Lu (Uni Florence): OOA-based landslide change detection  Also focused on multi-scale segmentation optimization Post-event image Landslide map Pre-event image OOA-based landslide Objective segmentation Change detection mapping (POF) Lu et al., 2011 Martha et al., 2011 Martha et al., 2010 (IEEE GSRL) (IEEE TGRS) (Geomorphology) 14 Image Mining 2013 – Barcelonette - 28 August 2013

Problem: object feature selection Hazard: Landslide work  Andre Stumpf: classification parameter and threshold selection  How to chose from hundreds of object features and the best threshold? Random Forest method (data mining/active learning based on samples)  15 Image Mining 2013 – Barcelonette - 28 August 2013

Hazard: Landslide work  Tested on air- and spaceborne data of 4 different sites  Accuracies of 73-87% OOA-based landslide Objective segmentation Change detection Objective parameter mapping (POF) Lu et al., 2011 selection Martha et al., 2011 Martha et al., 2010 (IEEE GSRL) Stumpf & Kerle, 2011 (IEEE TGRS) (Geomorphology) (RSE) 16 Image Mining 2013 – Barcelonette - 28 August 2013

Problem: limits of pan-chromatic data Hazard: Landslide work  Tapas Martha: OOA-based landslide detection based only on pan-chromatic data  Again use of POF  Focus on texture measures, segment refinement  Time-series analysis Landslides 17 Image Mining 2013 – Barcelonette - 28 August 2013

Hazard: Landslide work Objective segmentation Change detection … Objective parameter Use of pan- (POF) Lu et al., 2011 selection chromatic data Martha et al., 2011 (IEEE GSRL) Stumpf & Kerle, 2011 Martha et al, 2012 (IEEE TGRS) (RSE) (ISPRS) 18 Image Mining 2013 – Barcelonette - 28 August 2013

Problem: work with lidar data Hazard: Landslide work with LiDAR data  So far all work focused on optical data  Miet Van Den Eeckhaut: detection of forested landslides in single-pule LiDAR data  No use of additional optical data → focus on geomorphometry  Area in Flanders, Belgium; > 200 old deep-seated and shallow slides  Almost impossible to detect in optical data (Elevation exaggeration x1; @Google Earth) Complex slide Rotational slide 19 Image Mining 2013 – Barcelonette - 28 August 2013

Hazard: Landslide work with LiDAR data  Procedure:  Creation of LiDAR derivatives  Multiple segmentation based on POF  Detection of main scarp  Downslope growing using evidence from side and base scarp, as well as interior  Good detection of deep slides (71% of main scarps, >50% of associated landslide body 20 Image Mining 2013 – Barcelonette - 28 August 2013

Hazard: Landslide work with LiDAR data  Promising results given the challenging terrain OOA and LiDAR Van Den Eeckhaut et al., 2012 (Geomorphology) 21 Image Mining 2013 – Barcelonette - 28 August 2013

Hazard: Erosion detection Shruthi Rajesh: use of high-resolution satellite data to map gully erosion   Similar approaches to what we developed for landslides (directional texture, etc.)  Removal of false positives was challenging 22 Image Mining 2013 – Barcelonette - 28 August 2013

Problem: linear element detection Hazard: Erosion detection Gully erosion detection Shruthi et al., 2011 (Geomorphology) 23 Image Mining 2013 – Barcelonette - 28 August 2013

Problem: change detection of lines Hazard: Erosion detection Change detection for gully systems (2001-2009)  Gully detection with Random Forests Gully system change detection Shruthi et al., in review (Geomorphology) Shruthi et al., in press (Catena) 24 Image Mining 2013 – Barcelonette - 28 August 2013

Problem: building extraction Other risk aspects – Elements at risk Janak Joshi: Problem - building extraction from optical satellite data   Chicken & egg: we’d like to have a DEM/DSM, but photogrammetry is imperfect Z X, Y  Solution:  Create an (imperfect) DEM/DSM  Use in OOA (distinguish buildings from similar looking low features)  Use the extracted buildings to correct the DEM/DSM 25 Image Mining 2013 – Barcelonette - 28 August 2013

Other risk aspects – Elements at risk Initial DSM Geoeye image OOA-derived buildings Assignment of height Evident errors Corrected DSM 26 Image Mining 2013 – Barcelonette - 28 August 2013

Other risk aspects – Elements at risk Improved DSM, useful for example for flood modeling  27 Image Mining 2013 – Barcelonette - 28 August 2013