Identifying pure urban image spectra using a learning urban image spectral archive (LUISA)

Transcription

1 Identifying pure urban image spectra using a learning urban image spectral archive (LUISA) Marianne Jilge, Uta Heiden, Martin Habermeyer, André Mende, Carsten Juergens

2 Introduction Urban surface materials are essential for several urban studies (e.g. urban microclimate, surface runoff) Source: Source: large/ utility-1.jpg 2

3 Chart 3 LUISA June LUISA June 2016.de Introduction Chart Coating 3 - pure pixels in urban areas - Urban complexity Pure pixel = pixel composed of only one surface material class spectrally identifiable Trends Usage of material TriD72MwW1I/AAAAAAAAAQE/5Gkw M6eGpZ0/s1600/IMG_0138.JPG files/styles/blog_full/public/14.jpg?itok=pvot9jlc Aging Spatial Resolution mixed pixel pure pixel 3

4 Pure urban surface materials Automated endmember extraction Comparison with spectral libraries Source: Source: LUISA Learning urban image spectral archive Image-based identification of urban surface materials Generic urban spectral archive consideration of incompletness Usage of extracted pure material spectra for further applications 4

5 High resolution hyperspectral imagery 5 LUISA - Concept Learning urban image spectral archive (LUISA) LUISA-A Spectral archive Generically structured Continuously expandable Universally applicable due to: - sensor-based resampling - spectral redundancy reduction (ISODATA) LUISA-T Automatic identification of pure pixels known pure pixels Pure pixel Thresholding unknown pure pixels Dissimilarity Temporary Spec. Lib. Mixed pixel removal Based on >5,200 image spectra and different areas Focus on artifical materials mask of pure labeled pixels Intra-/Inter-Class- Homogeneity mask of categorized unknown pure pixels Result: scene-based spectral library of pure material spectra

6 LUISA-A Hierarchy * Arnold et al. (2013): The EAGLE concept - A vision of a future European Land Monitoring Framework 6

7 Statistically dominant similarity class LUISA-T: Pure Pixels known pure pixels Pure pixel Thresholding Measure SAM (Spectral Angle Mapper) SID (Spectral Information Differgence) SID-SAM SCM (Spectral Correlation Measure) SCA (Spectral Correlation Angle) SID-SCA JMD (Jeffries-Matusita Distance) LUISA-A n comparisons / pixel n similarity values / pixel Value Class Relationship Asphalt [1] Concrete [3] Water [5] Selection of similarity measure Pixel-wise comparison with LUISA-A spectra Ranking of similarity values Value Class Relationship Asphalt [1] Concrete [3] Asphalt [2] Concrete [1] Concrete [6] Concrete [4] Bitumen [5] Concrete [2] Asphalt [5] Bitumen [2] Tar Paper [1] Determination of statistically dominant similarity class (pre-classification) JMD-SAM 7

8 LUISA-T: Pure Pixels known pure pixels Pure pixel Thresholding Automated pure pixel thresholding Histogram of statistically dominant similarity values Second derivative Known pure pixels Separation of artifical and natural pixels Histogram statistics of statistically dominant similarity values Median filtered histogram Thresholding point: highest change between similarity and dissimilarity local maxima of second derivative mask of pure labeled pixels 8

9 Number of pixels LUISA-T: Unknown Pure Pixels unknown pure pixels Potentially pure unknown pixels Dissimilarity Temporary Spec. Lib. Mixed pixel removal Intra-/Inter- Class- Homogeneity Histogram of statistically dominant similarity values value 0 1 Temporary Spectral Library n comparisons / pixel Potentially unknown pure pixels Value Class Relationship Asphalt [1] Concrete [3] Water [5] better (lower) similarity values than first similarity analysis? Extraction of unsimilar spectra (dissimilarity analysis) Integrate outliers (similarity analysis) Spectral mixtures or pure unknown pixels 9

10 LUISA-T: Unknown Pure Pixels unknown pure pixels Dissimilarity Mixed pixel removal and identification of material classes Potentially unknown pure pixels Outlier erosion Border erosion Temporary Spec. Lib. Mixed pixel removal Intra-/Inter- Class- Homogeneity 4 pixel neighborhood Unknown pure pixel Spatial clusters Spectral intra-class homogeneity Spectral inter-class homogeneity mask of categorized unknown pure pixels Mixed pixels typically occur at/in object borders and single pixels (too small objects) Material classes are spectrally homogeneous and spatially independent 10

11 Applying LUISA Ludwigsburg, Germany Sensor: HyMap Date of acquisition: August 4th, 2010 Spatial resolution: 4m Number of spectral bands: 110 Band combination: R=1671nm, G=727nm, B=544nm LUISA Pre-processing: ias filtered (Rogge & Rivard, 2010) Rogge, D. and Rivard, B., Iterative spatial filtering for reducing intra-class spectral variability and noise, Proc. Hyperspectral Image and Signal Processing: Evolution in Remote Sensing (WHISPERS), 1-4, (2010). 11

12 Applying LUISA Artifical pure materials: Measure: SID-SCA 11.8% pure urban pixels threshold: Legend material classes aluminium concrete on roofs asphalt vegetation bitumen roofing tiles unknown abiotic material class 1 polyethylene PVC red loose chippings concrete on streets unknown abiotic material class 2 unknown or material mixture 12

13 Applying LUISA Natural pure materials: Measure: SID-SCA 21.01% pure natural pixels threshold: Legend material classes aluminium concrete on roofs asphalt vegetation bitumen roofing tiles unknown abiotic material class 1 polyethylene PVC red loose chippings concrete on streets unknown abiotic material class 2 unknown or material mixture 13

14 Applying LUISA Unknown artifical pure materials: 8 unknown material classes threshold (SAM): 0.1 Legend material classes aluminium concrete on roofs asphalt vegetation bitumen roofing tiles unknown abiotic material class 1 polyethylene PVC red loose chippings concrete on streets unknown abiotic material class 2 unknown or material mixture 14

15 Validating LUISA results Post-classification: Maximum likelihood classifier Applying extracted pure pixels as training data Validation based field data and expert knowledge Overall accuracy: 79.8% Kappa: 0.76 Legend material classes aluminium concrete on roofs asphalt vegetation bitumen roofing tiles unknown abiotic material class 1 polyethylene PVC red loose chippings concrete on streets unknown abiotic material class 2 unknown or material mixture 15

16 Conclusion and Outlook Automatic derivation of pure material spectra from an urban high resolution hyperspectral imagery Derivation of unknown material spectra Consideration of an incomplete spectral archive universally applicable Usage of extracted pure material spectra for further application Removal of remaining mixed pixel from mask of unknown pure pixels Consideration of albedo is crucial for mapping urban surface materials!! 16

17 Thank you for your attention!