2_Goodenough_IGARSS11_Final.ppt
- 1. Linear and Nonlinear Imaging Spectrometer Denoising Algorithms Assessed Through Chemistry Estimation David G. Goodenough 1,2 , Geoffrey S. Quinn 3 , Piper L. Gordon 2 , K. Olaf Niemann 3 and Hao Chen 1 1 Pacific Forestry Centre, Natural Resources Canada, Victoria, BC 2 Department of Computer Science, University of Victoria, Victoria, BC 3 Department of Geography, University of Victoria, Victoria, BC
- 2. Linear and Nonlinear Denoising Algorithms Assessed Through Chemistry Estimation Objective: To compare linear and non-linear methods of denoising hyperspectral data; do we always need non-linear methods? Data collection: Study area, sample collection, data/sensor characteristics Pre-processing: Orthorectification and radiometric calibration Processing: Contextual filter, spectral transformations, PLS regression, Chlorophyll-a and Nitrogen estimation Analysis: 30 x 30 m Plot-level 2 x 2 m Tree-level Conclusions
- 3. Data collection: The Greater Victoria Watershed District (GVWD) 14 plots, 140 trees
- 4. Acquisition date September 11, 2006 Spectral data Range: 395 - 2503nm 492 spectral bands Mean sampling interval: 2.37nm (VNIR <990nm) 6.30nm (SWIR>1001) Mean FWHM: 2.37nm (VNIR) 6.28 (SWIR) Spatial data 300 spatial pixels FOV: 22 ° IFOV: 0.076 ° Imaging rate: 40f/s Flight speed: 70m/s Along track sampling: 1.75m Flight altitude: 1500m 2m resolution Data collection: AISA Hyperspectral Data Acquisition
- 5. Sensor characteristics Discrete return LIDAR system 1064 nm FOV: 20 ° Footprint: ~25 cm (variable) Pulse rate: 100+ Khz Scan rate: 15 to 30 Hz Flight speed: 70 m/s Flight altitude: 1500m Posting density: ~1.2/m 2 Data Applanix 410 IMU/DGPS system First and last return x, y, z positions Range accuracy: 5 to 10 cm Rasterized to 2m resolution corresponding to AISA data Canopy height, digital surface and bare earth models are derived Data collection: Lidar Data Acquisition Acquisition date Concurrent with AISA acquisition
- 6. Geometric distortions (non-uniform distance and direction) caused by platform altitude, attitude (roll, pitch and yaw) and surface relief Traditional DEM orthorectification at fine resolutions introduce significant errors in tree canopy positions Accurate positioning is vital for high resolution datasets and fine scale patterns and processes The lidar RBO (range based orthorectification), reduces misregistration issues caused by layover of the reflected surface. Atmospheric corrections performed by ATCOR-4 (airborne) software applying sensor and atmospheric parameters to sample MODTRAN LUT and provide correction factors Empirical line calibration performed to reduce residual errors Data pre-processing: Radiometric and Geometric Correction AISA (B,G,R: 460,550,640nm) draped over LIDAR DSM
- 7. Nonlinearity of Hyperspectral Hyperspectral data is non-linear Minimum Noise Fraction (MNF) Popular linear noise removal technique Non-linear Local Geometric Projection Algorithm (NL-LGP) Will it outperform MNF denoising for foliar chemistry prediction? T. Han and D. G. Goodenough, "Investigation of Nonlinearity in Hyperspectral Imagery Using Surrogate Data Methods," Geoscience and Remote Sensing, IEEE Transactions on, vol. 46, pp. 2840-2847, 2008.
- 8. Denoising: Linear and Nonlinear AISA image 180 m x 170 m area True colour RGB: 1736, 1303, 1089nm Difference Images Inverse MNF denoised NL-LGP denoised NL-LGP - Reflectance Reflectance - MNF
- 9. NL-LGP Algorithm Construct state vectors in phase space Specify the neighbourhood of these state vectors Find projection directions Project the state vectors on these directions, reducing noise D. G. Goodenough, H. Tian, B. Moa, K. Lang, C. Hao, A. Dhaliwal, and A. Richardson, "A framework for efficiently parallelizing nonlinear noise reduction algorithm," in Geoscience and Remote Sensing Symposium (IGARSS), 2010 IEEE International , pp. 2182-2185.
- 10. Minimum Noise Fraction Estimates noise in the data and in a Principal Components Analysis ( PCA ) of the noise covariance matrix Noise whitening models the noise in the data as having unit variance and being spectrally uncorrelated A second PCA is taken Resulting MNF eigenvectors are ordered from highest to lowest signal to noise ratio (noise variance divided by total variance)
- 11. Plot-Level Chemistry Comparison Process AISA 30m data AISA 2m data MNF denoised data NL-LGP denoised data Averaging Inverse MNF denoising NL-LGP denoising Reflectance chemistry predictions MNF denoised chemistry predictions NL-LGP denoised chemistry predictions Chemistry ground data Partial Least Squares (PLS) Regression PLS Regression PLS Regression
- 12. Spectral Transformation for Comparing Chemistry Predictions Mean R 2 values for the transformation types are output by the PLS program Large standard deviations, overlapping between original reflectance, MNF and NL-LGP denoised 2 nd derivative (2 points left) has one of the highest mean R-squared values The most accurate predictions from PLS regression are output for each transformation type 2 nd derivative (2 points left) gave best prediction for all 3 spectra types and both Nitrogen and Chlorophyll-a chemistry
- 13. Plot-Level Average R-squared Values for Nitrogen
- 14. Plot-Level Non-current Nitrogen (% dry weight)
- 15. PLS Plot-Level Chlorophyll-a (μg/mg)
- 16. Moving from Plot-Level to Tree-Level Original reflectance predicts chemistry with greater accuracy than denoised reflectance Averaging from 2 x 2 m pixels to 30 x 30 m pixels Preprocessing of the data ( orthorectification and radiometric calibration) To find if there is non-linear noise at the 2 m level (tree-level) the process is repeated with original, non-averaged AISA 2 m data
- 17. Tree-Level Chemistry Comparison Process AISA 2m data MNF denoised data NL-LGP denoised data Inverse MNF denoising NL-LGP denoising Reflectance chemistry predictions MNF denoised chemistry predictions NL-LGP denoised chemistry predictions Chemistry ground data Partial Least Squares (PLS) Regression PLS Regression PLS Regression
- 18. Tree-Level Chemical Analysis Spectra were extracted from the positions of each tree in the plot data (2m by 2m pixels) Chemistry predictions were generated for the ten trees in each of the 14 plots, against the averaged chemistry measurement for their plot 2 nd derivative of reflectance (2 points left) gave the best R 2 values and was used for the chemistry predictions
- 19. Tree-Level Chemistry Comparison 14 Plots 140 Trees Predicted Chemistry for each of… MNF denoised NL-LGP denoised AISA 2m reflectance Averaged Measured Chemistry vs
- 20. PLS Tree-Level Non-current Nitrogen (% dry weight)
- 21. PLS Tree-Level Chlorophyll-a (μg/mg)
- 22. Conclusions: Linear and Non-Linear Denoising Algorithms For plot-level applications, denoising is not necessary The averaging process is effective for removing noise For tree-level applications, use of a non-linear denoising method is better for mapping chemistry Nitrogen Non-Linear 0.811 ± 0.047 MNF 0.679 ± 0.061 Original Reflectance 0.775 ± 0.051 Chlorophyll Non-Linear 0.818 ± 0.054 MNF 0.691 ± 0.061 Original Reflectance 0.758 ± 0.054
- 23. Conclusions: Linear and Non-Linear Denoising Algorithms MNF does not improve chemistry predictions, further supporting the non-linearity of hyperspectral data The application of PLS regression to forest chemistry mapping remains our most reliable method for chemistry estimation R 2 of ~0.9 for plot-level R 2 of ~0.8 for tree-level
- 24. We thank: The University of Victoria for its support. Natural Resources Canada (NRCan), the Canadian Space Agency (CSA), and Natural Sciences and Engineering Research Council of Canada (NSERC) (DGG) for their support. The Victoria Capital Regional District Watershed Protection Division for its logistical support. The audience for their attention. Acknowledgements: Hyperspectral applications for forestry
Editor's Notes
- #5: Note there are actually 296 spatial pixels with data. Spatial pixel number 1 and 4 contain an unresponsive FODIS (fiber optic downwelling irradiance system) data and 2 and 3 contain dark current data.