Small-footprint full-waveform airborne LiDAR for habitat assessment in the ChangeHabitats2 project Werner Mücke, András Zlinszky, Sharif Hasan, Martin Pfennigbauer, Hermann Heilmeier and Norbert Pfeifer werner.muecke@geo.tuwien.ac.at Research Groups Photogrammetry & Remote Sensing Department of Geodesy and Geoinformation Vienna University of Technology geo.tuwien.ac.at
Motivation (c) European Commission 2011: ~760 000 km² ~ 18% of EU landmass Netherlands, November 2013 W. Mücke 2
Motivation Natura 2000 habitat quality assessment Carried out by ecologists / biologists 760k km² experts in landscape character assessment AND in-situ, manually and terrestrially Obligatory frequent monitoring Reporting periods of 6 years More realistically it s 10 years! How can we use remote sensing to support this process? Netherlands, November 2013 W. Mücke 3
Motivation?? LiDAR data become increasingly available in better and better quality with wider and wider coverage There is more to them than?? / Legend LIDAR coverage of Europe? none/no data no info full scan complete partial LIDAR coverage just a DSM and a DTM ;-) full scan in progress Netherlands, November 2013 W. Mücke 4
CH2 approach and selected applications Habitat quality can be defined by three main parameters: Species Structure Disturbance Assess various metrics based on LiDAR data Some describe parameters directly, others indirectly (proxies) 3 applications and derivation of relevant metrics 1. Classification of grassland types 2. Identification of dead wood in forests 3. Detection of linear features in landscapes Netherlands, November 2013 W. Mücke 5
Grassland types: Aims Not vegetation Wet high Open soil Wet meadow Abandoned Fringe Dry meadow Lowland hay meadow Meadow - like Netherlands, November 2013 W. Mücke 6
Grassland types: Study site and data Field data May + Sept 2012, April + May 2013 mapped literally everything they saw and could parameters related to species, structure or disturbance LiDAR data July 2011 (leaf-on) + March 2012 (leaf-off) RIEGL LMS-Q680i point density: last echoes: 12,8 / m² Height accuracy: 3.1 cm Swath discrepencies Netherlands, November 2013 W. Mücke 7
Training data Grassland types: Method(s) Classifier settings Learning process Validation data LiDAR input / primary layers 0,5 * 0,5 m² rasters of... Echo Width Reflectance Sigma Z Openness LiDAR metric Correlates with vegetation... biomass, penetration water content (dryness) surface roughness, penetration surface pattern Subset of layers Expert validation Trained classifiers Classified vegetation maps Quality control ndsm Canopy height Netherlands, November 2013 W. Mücke 8
Grassland types: Method(s) 0,5 * 0,5 m² rasters of... Echo Width Training data LiDAR metric Correlates with vegetation... Classifier settings biomass, penetration Learning process Validation data LiDAR Reflectance input / primary layers Sigma Z Openness ndsm water content (dryness) surface roughness, penetration surface pattern Subset of layers Canopy height Expert validation Trained classifiers Classified vegetation maps Quality control Netherlands, November 2013 W. Mücke 9
Training data Grassland types: Method(s) Classifier settings Learning process Validation data LiDAR input / primary layers 0,5 * 0,5 m² rasters of... Echo Width Reflectance Sigma Z Openness ndsm LiDAR metric Correlates with vegetation... surface roughness, penetration water content (wet- or dryness) surface roughness, penetration surface pattern Canopy height Subsets of layers Expert validation Trained classifiers Classified vegetation maps Quality control Netherlands, November 2013 W. Mücke 10
Grassland types: Example LiDAR subsets in study site 50 Meters Netherlands, November 2013 W. Mücke 11
Grassland types: Results No data / masked 50 Meters Netherlands, November 2013 W. Mücke 12
Grassland types: Results No data / masked 50 Meters Adding class probability Netherlands, November 2013 W. Mücke 13
Grassland types: Results and evaluation Classified as Reference class not open low. hay meadow dry wet user's acc. vegetation soil shrub fringe abandoned mea. like Lawn wet meadow meadow high totals [%] not vegetation 19941 800 61 448 568 11 5 65 0 25 2 21926 90,9 open soil 2 1849 21 257 129 565 250 0 1 99 0 3173 58,3 Automatic shrub 11 3 2040 591 129 0 130 0 0 1 0 2905 70,2 fringe 52 168 97 1814 436 203 91 2 18 109 504 3494 51,9 Reproducible abandoned 0 108 96 336 2042 129 54 0 39 443 118 3365 60,7 lowland hay meadow 0 942 78 429 218 1776 29 172 2 725 392 4763 37,3 Robust meadow like 55 105 38 678 77 204 295 27 0 154 116 1749 16,9 artificial lawn 1166 164 3 17 34 1 33 2111 0 0 0 3529 59,8 wet meadow 0 1 0 231 13 41 157 0 336 44 830 1653 20,3 dry meadow 11 555 234 15 210 109 238 0 69 1321 36 2798 47,2 wet high 0 180 0 197 370 56 1 0 276 2 3731 4813 77,5 totals 21238 4875 2668 5013 4226 3095 1283 2377 741 2923 5729 54168 producer s acc. [%] 93,9 37,9 76,5 36,2 48,3 57,4 23,0 88,8 45,3 45,2 65,1 68,8 Netherlands, November 2013 W. Mücke 14
Deadwood detection: Aims Netherlands, November 2013 W. Mücke 15
Deadwood detection: Study site and data ALS data Spring 2012 (March 28) RIEGL LMS-Q680i point density: all echoes: 29,4 / m² last echoes: 10,9 / m² Height accuracy: 2.9 cm Strip discrepancies Netherlands, November 2013 W. Mücke 16
Deadwood detection: Study site and data Field data Collected in Spring, Summer and Winter 2012 GNSS and tape measurements of Downed stems Snags and fine woody debris Photographs Netherlands, November 2013 W. Mücke 17
Deadwood detection: Method Stepwise process Filtering of point cloud Morphological image processing Raster map algebra Final result: Regions (i.e. 2D outlines) of downed stems 18
Deadwood detection: Results and evaluation Classification based on Shape of object /Area / perimeter) Surface roughness (cf. Ew var ) Netherlands, November 2013 W. Mücke 19
Deadwood detection: Results and evaluation W. Mücke, B. Deák, A. Schroiff, M. Hollaus & N. Pfeifer: Detection of fallen trees in forested areas using full-waveform airborne laser scanning data. Canadian Journal of Remote Sensing (accepted, already available online) Completeness = 75, 6 % Correctness = 89, 9 % Netherlands, November 2013 W. Mücke 20
Linear feature detection: Aims Find indicators for human influence on landscapes Netherlands, November 2013 W. Mücke 21
Linear feature detection: Method(s) Combination of various LiDAR derivatives DTM, ndsm, slope, curvature, openness, reflectance, echo width, echo count Raster map algebra Classification (rule-based, SVM) Morphological image processing Hough transform Netherlands, November 2013 W. Mücke 22
Linear feature detection: Examples Powerline detection 1. Aerial image 2. Orienteering map 3. Classified powerline pixels 4. Powerline delineated Netherlands, November 2013 W. Mücke 23
Linear feature detection: Examples Asphalt roads and other linear features Netherlands, November 2013 W. Mücke 24
Conclusions Landscape assessment w.r.t. N2000 is a complex task Manual field work and expert knowledge is and will always be necessary Remote sensing can support the process Save manual (field) work save time save money with robust, reproducible and automatic methods Identification: find hot spots, areas of significant change Quantification: provide objective, area wide estimations CH2 aims at supporting the otherwise cumbersome assessment process in every possible way! Netherlands, November 2013 W. Mücke 25
Acknowledgements The ChangeHabitats2 Team especially Anke Schroiff, Balázs Deák, Agnés Vari, Adam Kania, Hermann Heilmeier RIEGL Laser Measurment Systems GmbH kindly provided their company airplane, some laser scanners and working hours Some pictures & figures I borrowed from http://ec.europa.eu/index_de.htm http://www.proleipzig.eu/html/pressearchiv.html www.skiresort.de Google Maps - Panoramio Netherlands, November 2013 W. Mücke 26
The obligatory extra slides... ;-) In the following some additional information on Grassland type estimation Dead wood identification Linear feature detection Netherlands, November 2013 W. Mücke 27
Grassland types: Example LiDAR subsets in study site Reflectance Echo width Openness sigmaz ndsm 50 Meters Netherlands, November 2013 W. Mücke 28
Deadwood detection: Method Ohne EW-Filter Mit EW-Filter Stepwise process Filtering of point cloud Morphological image processing Raster map algebra Final result: Regions of downed stems Netherlands, November 2013 W. Mücke 30
Deadwood detection: Results and evaluation Reference stem was marked as found, if correspondence with automatically derived outline was bigger than 75% partly found, if correspondence with outline was less than 75% not found, if no corresponding outline was found Netherlands, November 2013 W. Mücke 31
Deadwood detection: Results and evaluation Completeness = TP ( TP + FN ) Correctness = TP ( TP + FP ) With a TP is True Positive (TP): stem was found and is in reference If it was found and decay state was 1,2 or 3 False Positive (FP): found, but not in reference If partly found and decay state was 2 or 3 False Negative ELMF (FN): Amsterdam, not found, The but in reference Netherlands, November 2013 W. Mücke 32
Deadwood detection: Comparison of discrete and FWF ALS II I III TP Completeness = ( TP + FN ) TP Correctness = ( TP + FP ) Result from discrete ALS includes more FPs! With True Positive (TP): stem was found and is in reference False Positive (FP): found, but not in reference False Negative (FN): not found, but in reference Netherlands, November 2013 W. Mücke 33
Deadwood detection: Conclusions Well-preserved stems are found reliably Also if they are covered by fine woody debris or vital vegetation (31 of 40 pc.) Decay state and diameter have major influence < 10 cm was impossible DTM calculation method should Minimize influence of low vegetation BUT not at the same time eliminate small scale topography Influences of sensor characteristics and flight parameters Adaption of thresholds Netherlands, November 2013 W. Mücke 34