HYPERSPECTRAL RESOLUTION ENHANCEMENT USING MULTISENSOR IMAGE DATA - PowerPoint PPT Presentation

HYPERSPECTRAL RESOLUTION ENHANCEMENT USING MULTISENSOR IMAGE DATA J. Bieniarz, D. Cerra, X. X. Zhu, R. Müller, P. Reinartz German Aerospace Center (DLR), Remote Sensing Technology Institute (IMF), Oberpfaffenhofen, 82234 Wessling, Germany. jakub.bieniarz@dlr.de +49 8153 28-2790

www.DLR.de • Chart 7 Unmixing Based Resolution Enhancement

www.DLR.de • Chart 8 Unmixing Based Resolution Enhancement

www.DLR.de • Chart 9 Highly Mixed Scenario ? ?

www.DLR.de • Chart 10 Multi-Look Joint Sparsity Fusion (MLJSF) - Unmixing based image fusion - Use of external (pre-defined) spectral libraries containing pure endmemebers - Jointly sparse unmixing concept – spatial spectral imnixing

www.DLR.de • Chart 11 Sparse Spectral Unmixing – The Concept y = A x = Use of large libraries • Aster Spectral Library from JPL, USGS, JHU with 2463 spectra • DLR Spectral Archive with 1609 spectra Requirements: significant endmembers • Higly accurate Atmospheric Correction -

www.DLR.de • Chart 12 Hyperspectral Image Resolution Enhancement Based on Jointly Sparse Spectral Unmixing HSI Sparse Spectral MSI unmixing dictionary (window based) of HSI Resample Joint Sparse Select endmembers unmixing Endmembers to MSI SRF of MSI (window based) High resolution HSI (window based)

www.DLR.de • Chart 13 Hyperspectral Image Resolution Enhancement Based on Jointly Sparse Spectral Unmixing HSI . . Sparse . . Spectral MSI . unmixing . dictionary (2f x 2f window) . of HSI . . . Resample Jointly sparse Select . . endmembers unmixing Endmembers to MSI SRF of MSI (window based) From a window select active endmembers and construct a new pruned High resolution dictionary for this region. HSI (af x bf Window)

www.DLR.de • Chart 14 Hyperspectral Image Resolution Enhancement Based on Jointly Sparse Spectral Unmixing HSI Sparse Spectral MSI unmixing dictionary (window based) of HSI Resample Joint Sparse Select endmembers unmixing Endmembers to MSI SRF of MSI (2 x 2 window) Unmix the window of the MSI using the Joinr Sparsity Model (JSM) High resolution HSI (window based)

www.DLR.de • Chart 15 EnMAP + World_View2 + Thanks to Karl Segel

WV2 EnMAP Reference MLJSF Standard method

WV2 EnMAP Reference MLJSF Standard method

www.DLR.de • Chart 19 Experiments – spatial resolution HYPERION ● 30m spatial resolution ● 220 spectral bands (152 after band reduction) ● 10nm spectral resolution ● 0.4-2.5 spectral range

www.DLR.de • Chart 20 Experiments – spatial resolution LANDSAT 7 (simulated from AISA) ● 3m spatial resolution ● 6 spectral bands ● 0.4-2.4 spectral range

www.DLR.de • Chart 21 Experiments – spectral resolution HYPERION ● 30m spatial resolution ● 220 spectral bands (152 after band reduction) ● 10nm spectral resolution ● 0.4-2.5 spectral range Landsat 7 SRF Spectrum LANDSAT 7 (simulated from AISA) ● 3m spatial resolution ● 6 spectral bands ● 0.4-2.4 spectral range

www.DLR.de • Chart 22 Results MLJSF ● 3m spatial resolution ● 152 spectral bands ● 10nm spectral resolution ● 0.4-2.5 spectral range

Evaluation – 2 ways ● Performance measure ● Quantitative comparison ● Application example ● SAM Classification

www.DLR.de • Chart 24 Quantitative analisys

www.DLR.de • Chart 25 Quantitative analisys

www.DLR.de • Chart 26 Application to classification HYPERION ● 30m resolution ● OA: 42% ● Kappa:0.36 Compared to classification results of AISA original image

www.DLR.de • Chart 27 Application to classification MLJSF ● 3m resolution ● OA: 79% increase of 37% ● Kappa:0.76 increase of 0.40

www.DLR.de • Chart 28 Application to classification HYPERION ● 30m resolution ● OA: 42% ● Kappa:0.36 Compared to classification results of AISA original image

www.DLR.de • Chart 29 Application to classification CNMF ● 3m resolution ● OA: 64% increase of 22% ● Kappa:0.59 increase of 0.23

www.DLR.de • Chart 30 Conclusions ● MLJSF exploits the sparsity in hyperspectral mixtures and the HR spatial information in the multispectral image. ● MLJSF achieves ● Comparable quantitative metrics to the state-of-the-art methods, even when using external libraries containing pure spectra. ● Significantly better performance when aiming at applications like classification. ● Future work: Robustness analysis against co- registration error, different conditions of the data acquisition and missing materials in the library etc.

HYPERSPECTRAL RESOLUTION ENHANCEMENT USING MULTISENSOR IMAGE DATA J. Bieniarz, D. Cerra, X. X. Zhu, R. Müller, P. Reinartz German Aerospace Center (DLR), Remote Sensing Technology Institute (IMF), Oberpfaffenhofen, 82234 Wessling, Germany. jakub.bieniarz@dlr.de +49 8153 28-2790

www.DLR.de • Chart 32 Spectral Resolution

www.DLR.de • Chart 33 Spectral Resolution

www.DLR.de • Chart 34 Highly mixed scenario

www.DLR.de • Chart 35 Hyperspectral Image Resolution Enhancement Based on Jointly Sparse Spectral Unmixing HSI . Sparse . Spectral MSI . unmixing dictionary (2f x 2f window) of HSI . . Resample Jointly sparse Select . endmembers unmixing Endmembers to MSI SRF of MSI (2 x 2 window) 1) Unmix the whole HSI (pixel-wise) using the sparse spectral unmixing method and High resolution a priori given spectral dictionary. HSI (af x bf Window)

www.DLR.de • Chart 36 Hyperspectral Image Resolution Enhancement Based on Jointly Sparse Spectral Unmixing HSI Sparse Spectral MSI unmixing dictionary (window based) of HSI Resample Joint Sparse Select endmembers unmixing Endmembers to MSI SRF of MSI (2 x 2 window) 3) Resample endmember spectra to the MSI sensor spectral resolution using the spectral response function (SRF) High resolution HSI (window based)

www.DLR.de • Chart 37 Hyperspectral Image Resolution Enhancement Based on Jointly Sparse Spectral Unmixing HSI . . Sparse . Spectral MSI unmixing . dictionary (window based) . of HSI . Resample Joint sparse Select endmembers unmixing Endmembers to MSI SRF of MSI (2 x 2 window) 5) reconstruct the high resolution HSI, with the resulting MSI abundances and the original HSI spectral dictionary using the High resolution LMM. HSI (window based)