DETECTING AND CORRECTING THE CURVATURE EFFECT (SMILE) IN HYPERION IMAGES BY USE OF MNF
Alon Dadon (1), Arnon Karnieli (1), Eyal Ben-Dor (2) t
The Remote Sensing Laboratory, Jacob Blaustein Institutes for Desert Research, Ben-Gurion University of the Negev, Sede-Boker Campus 84990,, Israel, Email: dadonalo@ bgu.ac.il ,
[email protected] (2) The Remote Sensing and GIS laboratory, Department of Geography and Human Environment, Tel-Aviv University, Israel, Email:
[email protected] ABSTRACT Various pushbroom imaging spectroscopy[1] systems experience low-frequency array effects that are often referred to as spectral curvature effect or smile/ frown interferences. The smile effect may render spectral absorption positions, leading to inadequate atmospheric correction and a noisy reflectance product; as a result, the thematic spectral based products may be erroneous e.g., thematic misclassification of the data. A method for the correction of spectral curvature effects (smile) in Hyperion (EO-1) images, termed “Trend Line Smile Correction” (TLSC), is presented. TLSC operates a precise modification to the MNF-1 band, based on the evaluation of the smile effect according to absorption features of well-mixed gases. The methodology was tested on four different Hyperion scenes and consistently outperformed other tested methods by up to 9 times. 1.
INTRODUCTION
Hyperion hyperspectral sensor onboard the EarthObserving-1 (EO-1) spacecraft is a pushbroom imaging spectrometer [2-4]. The detector elements in a pushbroom system are arranged on a rectilinear grid, therefore dispersing the slit onto the straight rows of detector elements leads to spatial misalignment of the wavelength and bandwidth, known as spectral curvature effect or smile/ frown interferences [5, 6]. Smile may also be a product of aberrations in the collimator and imaging optics [5]. Throughout the Hyperion operations, shifts of less than 1 nm have been observed in the SWIR wavelengths and 2.6 – 4.25 nm shifts in the VNIR detector wavelengths. The motivation for developing a smile correction methodology came from an unsuccessful classification scheme for geology applications. Where the spectral Angle Mapper (SAM) supervised classification method [7, 8] was applied to an Hyperion reflectance image of Dana Geology Reserve in southern Jordan. The main advantage of SAM is that it is relatively insensitive to illumination
_____________________________________________________ Proc. ‘Hyperspectral 2010 Workshop’, Frascati, Italy, 17–19 March 2010 (ESA SP-683, May 2010)
and albedo effects [9]. The final output product was a thematic geology map (Fig. 1).
Figure 1. spectral Angle Mapper (SAM) classification of Hyperion image of Dana Geology Park, Jordan. The SAM classification of Hyperion (Fig. 1) distinguished between most of the different rock types, excluding the right hand side of the image, for which only one class was assigned. This is ascribed to the smile malfunction appearing in Hyperion. 2.
PREVIOUS METHODS
Hyperion pre-launch smile estimates are inapplicable as the smile changes with time [10]. On the other hand, MNF-1 commonly used as a smile indicator does not fully represent the smile effect [5]. therefore, previous attempts to implement corrections to the MNF, such as column mean corrections (as in global destriping), fitting low-order polynomials as in “cross track illumination”, or even removing MNF-1 from the data, were ineffective and resulted in significant distortion of spectral and spatial data [5, 11]. One of the available techniques, implemented prior atmospheric correction, is the “cross-track illumination correction” module [12], provided within the ENVI software [13] used to remove cross track illumination variations. These may be due to vignetting effects, instrument scanning, or other non-uniform illumination effects often observed in pushbroom instruments. As this is the
only available method to deal with pushbroom affects, it is a necessary step in the correction methodology of Hyperion data and as such it is brought herein as a reference for comparison. Due to Hyperion malfunctions and low SNR, not much experience has been gained with the instrument and the smile effect remains a considerable challenge for the scientific community [2, 5, 14].
3.
viewing angles. TLSC includes a smile detection procedure that is performed by spectral comparison of the radiance image to the radiative transfer absorptions and by locating the minimal smile column in the data. The trend line smile correction modifies the MNF-1 band according to the calculated smile drift from the minimal smile column. The corrected MNF band is than inserted into the inverse MNF procedure, converting the coherent data back to radiance values. The final stage is a spectral comparison to MODTRAN for validation purposes.
TREND LINE SMILE CORRECTION (TLSC)
3.1 TLSC Rational Any smile correction methodology must consist of at least three stages: smile detection and quantification; correction of the detected cross-track variation; and evaluation of the correction performance in light of the user requirements. As noted, Hyperion pre-launch smile estimates are inapplicable, as the smile appearance varies from scene to scene [10]. On the other hand, the MNF-1 gradient does not fully represent the smile effect [5]. Therefore, the smile effect must be determined according to inherent per scene indicators. Previous methods attempting to correct smile, such as column mean corrections (as in global destriping), applying a polynomial correction to the gray scale gradient in MNF-1, or even removing MNF-1 from the data and fitting low-order polynomials as in “cross track illumination” [5, 11] are reported to result in distortion of the spectra [15]. Therefore, a precise modification to the MNF-1 band, based on the evaluation of the smile effect according to absorption features of well-mixed gases, is suggested. 3.2 TLSC Implementation The hypothesis behind this methodology takes into consideration the fact that MNF-1 consists of both spatial and spectral information. Therefore, its correction, according to exclusively spectrally derived parameters (e.g., atmospheric absorption features) will account for the smile effect in the data. Assuming that by isolating the spectral curvature effect component appearing in the first MNF and implementing a correction to it, the effect will be removed from the data during the inverse MNF stage with minimal influence on the spatial components that may contribute distortion of the spectra. The rationale of the described approach is to suggest an image-based spectral correction, for the main purpose of providing smile-corrected radiance data. In order to evaluate the validity of this imagebased approach, the procedure was carried out on 4 different images, acquired on different dates and
The TLSC method is based on the assumption that there is a partial correlation between data spectral nonuniformity due to smile and the gray scale gradient commonly appearing in the first MNF image (MNF-1). However, MNF-1 consists of both spatial and spectral information. Therefore, it is hypothesized that correcting MNF-1 according to exclusively spectrally derived parameters will account for the smile effect in the data. The correction methodology includes a smile detection procedure that is performed by spectral comparison of the radiance image to the radiative transfer absorptions and by locating the minimal smile column (MSC) in the data. A set of normalization factors calculated from the spectral derivative at the right hand side of the O2 absorption feature (760nm) were used as internal indicators to scale the initial MNF-1. A second order polynomial function, referred as Trend Line (TL), was fitted to the column mean derivative values of O2 absorption. The TL function was then scaled to fit the MNF-1 values. Comparison between derivatives calculated for the O2 absorptions of both Hyperion and MODTRAN [16] was used to locate of the minimal smile columns (MSCs). These columns, which are the ones list disturbed by smile, later served as reference factors in the correction methodology. Contrary to the accustomed polynomial fit, the adaptation was simply preformed by arithmetic addition and subtraction. Moreover the MSC in MNF space was used for fine-tuning. Selected MNF bands (including the corrected MNF-1) were used to reconstruct the image back to radiance space [17]. Similar spectral smile detection methodologies that utilize atmospheric absorption features were suggested [6, 10, 18-20] . The presented technique differs from these methods both in the detection and correction approaches.
4.
RESULTS
The methodology was tested on Four Hyperion images, covering the vicinity of the Arava Rift Valley, Israel and Jordan. Evaluation of the smile effect correction performance was carried out in three stages: (1) matching O2 and H2O absorption features across the
dataset with MODTRAN absorption features; (2) comparison with other commonly used methods; (3) comparing classification retrieval of selected targets with thematic mapping of the Dana Geology National Park area in southeastern Jordan . (b) Polinomial ENVI TLSC
0
60
60
(e)
0
120 180 Column
0
60
120 180 Column
60
120 180 Column
1 0.8 0
60
240
Non-corrected TLSC
0.7 0.6 0.5 0.4 0.3 0.2 0.1
0
60
(h)
240
120 180 Column
45 40 35 30 25 20 15 10 5 0
240
Deviation 0
240
In addition TLSC significantly reduces spikes near 940nm water vapor absorption region, implying that although TLSC does not deal directly with band-to-band smile differences, by using MNF the correction is applied to all bands (Fig.3).
0.9
(f)
45 40 35 30 25 20 15 10 5 0
120 180 Column
45 40 35 30 25 20 15 10 5 0
240
45 40 35 30 25 20 15 10 5 0
(g)
60
(d)
Deviation
Deviation
240
45 40 35 30 25 20 15 10 5 0 0
Deviation
120 180 Column
Deviation
Deviation
(c)
45 40 35 30 25 20 15 10 5 0
Reflectance
45 40 35 30 25 20 15 10 5 0
Deviation
Deviation
(a)
of O2 (Fig. 2 a-b) and H2O (Fig. 8 e-h) from that of the minimal smile column, for the tested correction methods. The polynomial and cross-track illumination correction methods exhibit higher values of deviation compared to those derived by the proposed TLSC procedure, except for one dataset (Kz), where values of the O2 absorption were similar (Fig. 2 a).
120 180 Column
240
45 40 35 30 25 20 15 10 5 0
0 426
626
826 Wavelength (µm)
1026
Figure 3. Example of a VNIR reflectance spectral curve of an image pixel (Ke, x:11, Y:331) before (thin line) and after TLSC correction (thick line). No empirical line or any other smoothing was applied. 0
60
120 180 Column
240
Figure 2. Standard deviation of O2 column mean derivative values (at 772 nm), from derivative values of the MODTRAN value across Kz, Kk, Ke and Ki scenes (a, b, c and d, respectively) subsequent to different smile corrections: polynomial fit correction (gray), the ENVI cross track illumination correction (thin black), and the Trend Line correction (thick black). Respectively, e-h present the results for H2O (at 823 nm). Derivative matching to the MSCs value (according to MODTRAN value) was preformed both for O2 (672 nm) and H2O (833 nm) absorption features representing the. TLSC reduced the cross- track variations in average by 8 times for both O2 (at 760 nm) and H2O (at 823 nm) absorption features and consistently surpassed other tested methods. Results of the TLSC were also compared to available methods commonly used in the preprocessing of Hyperion data i.e., the “cross-track illumination correction” module of ENVI, originally designed to remove the cross-track illumination variation of an image, and the polynomial smoothing correction, frequently applied to the MNF-1 band. Fig. 2 displays the cross-track deviation of derivative values
TLSC contributed significantly to the reliability of the reflectance product and consequently of the final classification map (Fig. 4).
(a)
(b)
Figure 4. VNIR-based classification for (a) non smile corrected reflectance image; (b) Trend Line Smile Corrected (TLSC) reflectance image. As can be seem in Fig. 2 the smile malfunction on the right side of the non corrected image (Fig. 4 a) has been resolved and all rock units appear correctly in the TLSC classification image (Fig. 4 b).
5.
CONCLUSIONS
Theoretically, the Hyperion data should provide improved geological information that currently cannot be obtained by any other space borne system. Nevertheless, the Hyperion data were found to be relatively noisy (average SNR of 40), in which significant smile effects were embedded that were difficult to handle by the common end-user. Therefore the smile effect must be determined according to inherent per scene indicators. The well-mixed gas absorption features are limited in number, however unlike surface features, they are relatively homogeneously spread across the image particularly in flat terrain areas. Hense thy were found to be useful as internal indicators for the smile effect. TLSC significantly reduces spikes near 940nm (water vapor absorption), implying that although TLSC does not deal directly with band-to-band smile differences, by using MNF the correction is applied to all bands. The success of TLSC in correctly revealing geological features over four particular scenes is encouraging. It is likely that correction of Hyperion data over other homogeneous arid areas would yield positive results. It is admitted, however, that until the proposed method be successfully tested on other environments and acquisition conditions it is not robust. The TLSC should be examined for the cases of (a) airborne imaging spectrometers; (b) low-to-medium smile; (c) rough terrain; and (d) in the presence of particular spatial patterns such as those of clouds or coastal lines. These all are subject to future research. Finally, it is our hope that the effectiveness of TLSC will improve the applicability of Hyperion data and possibly, of future pushbroom hyperspectral spaceborne missions for citizen science and social networks end-users. The authors wish to bring to more Israeli-Jordanian scientific cooperation for the evolvement of the research in the coming future.
6.
REFERENCES
1.
Asner GP, Heidebrecht KB. Imaging spectroscopy for desertification studies: Comparing AVIRIS and EO-1 Hyperion in Argentina drylands. IEEE Transactions on Geoscience and Remote Sensing 2003;41;1283-1296. Ungar SG, Pearlman JS, Mendenhall JA, Reuter D. Overview of the Earth Observing One (EO-1) mission. IEEE Transactions on Geoscience and Remote Sensing 2003;41;1149-1159. Ungar SG. Technologies for future Landsat missions. Photogrammetric engineering and remote sensing 1997;63;901- 905.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
Pearlman JS, Barry PS, Segal CC, Shepanski J, Beiso D, Carman SL. Hyperion, a space-based imaging spectrometer. IEEE Transactions on Geoscience and Remote Sensing 2003;41;11601173. Jupp DLB, Datt B, McVicar TR et al. Improving the Analysis of Hyperion Red-Edge Index from an Agricultural area. SPIE’s Third International AsiaPacific Remote Sensing symposium. Hangzhou, China, 2002;1-15. Neville RA, Sun L, Staenz K. Detection of spectral line curvature in imaging spectrometer data. SPIE, 2003: Algorithms and Technologies for Multispectral, Hyperspectral, and Ultraspectral Imagery IX., 2003;144-154 Boardman JW. Automated spectral unmixing of AVIRIS data using convex geometry concepts. Fourth JPL Airborne Geoscience Workshop: JPL, 1993;11 - 14. Chen J, Xiuping J, Wei Y, Matsushita B. Generalization of Subpixel Analysis for Hyperspectral Data With Flexibility in Spectral Similarity Measures. Geoscience and Remote Sensing, IEEE Transactions on 2009;47;21652171. Kruse FA, Lefkoff AB, Boardman JW et al. The Spectral Image-Processing System (Sips) Interactive Visualization and Analysis of Imaging Spectrometer Data. Remote Sensing of Environment 1993;44;145-163. Neville RA, Sun L, Staenz K. Spectral calibration of imaging spectrometers by atmospheric absorption feature matching Canadian Journal of Remote Sensing 2008; 34;29-42. Goodenough DG, Dyk A, Niemann O et al. Processing Hyperion and ALI for forest classification. IEEE Transactions on Geoscience and Remote Sensing 2003;41;1321-1331. Kennedy RE, Cohen WB, Takao G. Empirical methods to compensate for a view-angle-dependent brightness gradient in AVIRIS imagery. Remote Sensing of Environment 1997;62;277-291. Reserch System Inc. R. The Environmmentfor visualizing Images. ENVI User's Guide: Boulder, Co., 2008. Kerola DX, Bruegge CJ, Gross HN, Helmlinger MC. On-Orbit Calibration of the EO-1 Hyperion and Advanced Land Imager (ALI) Sensors Using the LED Spectrometer (LSpec) Automated Facility. Geoscience and Remote Sensing, IEEE Transactions on 2009;47;1244-1255. Datt B, McVicar TR, Van Niel TG, Jupp DLB, Pearlman JS. Preprocessing EO-1 Hyperion hyperspectral data to support the application of agricultural indexes. IEEE Transactions on Geoscience and Remote Sensing 2003;41;12461259.
16. Staenz K, Neville RA, Clavette S, Landry R, White HP. Retrieval of Surface Reflectance from Hyperion Radiance Data. IEEE Geoscience and remote sensing letters 2002;1;1419-1421. 17. Green AA, Berman M, Switzer P, Craig MD. A transformation for ordering multispectral data in terms of image quality with implications for noise removal. Geoscience and Remote Sensing, IEEE Transactions on 1988;26;65-74. 18. Barry PS, Shepanski J, Segal C. On-orbit spectral calibration verification of Hyperion. Geoscience and Remote Sensing Symposium, 2001. IGARSS '01. IEEE 2001 International, 2001;2535-2537 vol.2536. 19. Gao B, Montes MJ, Davis CO. Refinement of wavelength calibrations of hyperspectral imaging data using a spectrum-matching technique. Remote Sensing of Environment 2004;90;424-433. 20. Guanter L, Richter R, Moreno J. Spectral calibration of hyperspectral imagery using atmospheric absorption features. Applied Optics 2006;45;2360- 2370. 21. Richter R, Schlapfer D. Considerations on water vapor and surface reflectance retrievals for a spaceborne imaging spectrometer. IEEE Transactions on Geoscience and Remote Sensing 2008;47;1958- 1966.
