Advances in reducing radiometric miscalibration – application for hyperspectral push-broom sensors

Data takes of hyperspectral imagers are of increasing demand in Earth Observation related applications. As for other remote sensing techniques this requires precise pre-processing comprising of radiometric, spectral and geometric distortion reductions. One of these steps is radiometric scaling to transform recorded digital number to radiance. For this, laboratory assessed mathematical relations between required radiance and recorded digital number (gain) and shortterm measurements of dark current variations (offset) during operation are incorporated. Due to changes in the sensor system, which include thermal imbalance and mechanical stress gain and offset may vary over time. The result of this is visually perceptible as along-track striping noise after radiometric calibration. In this work, a new approach is presented that enables fast, highly precise and parameter-free destriping of uncorrelated striping noise that enhances the radiometric accuracy of hyperspectral push-broom data takes, and, hence, improves the outputs of succeeding applications. It is part of the existing ROME (Reduction of Miscalibration Effects) framework and is based on a noise-perpendicular gradient minimization technique. The performance was tested and compared to four state-of-the-art algorithms using artificially degraded hyperspectral whisk-broom scenes from a HyMAP campaign over Germany, two AISA scenes over Germany and two EO-1 Hyperion scenes over Namibia. Proposed approach clearly outperforms all other tested approaches even in low SNR scenarios like close to atmospheric absorption bands. On average a destriping accuracy of 99.75 % can be achieved having 3σ of only 1 % and, thus, it has been integrated into the state-of-the-art ROME framework that becomes a standard inside hyperspectral pre-processing chains. INTRODUCTION Remote sensing acquisitions often serve as a basis for geospatial applications. Especially imaging spectroscopy becomes more and more important because continuous pixel spectra enable the differentiation of surface cover materials. Modern sensors that utilize the push broom technology offer a broad applicability through the variable integration time but have to be carefully radiometrically calibrated. This is usually performed in special laboratories. Estimated calibration sets are then used to transform the raw hyperspectral data cube into physics based units such as at-sensor-radiance that generally builds the base for any application. If the characteristics of the detectors change over time and do not temporally coincide with last laboratory calibration, then visual perceptible along-track stripes occur in the acquisitions that aggravates succeeding applications. However, in this work a new approach is presented that significantly reduce stripes. It relates to human perception through its global gradient minimization principle and clearly outperforms inspected current state-of-the-art destriping approaches. It has been evaluated using data sets of three different hyperspectral sensors – HyMAP, AISA Dual and EO-1 Hyperion for which ground-truth was partly available. In the following the basic concepts and related results are briefly presented. EARSeL 34th Symposium Proceedings, 16-20 June 2014 4.18 © EARSeL and University of Warsaw, 2014, ISBN 978-83-63245-65-8, DOI: 10.12760/03-2014-09, Zagajewski B., Kycko M., Reuter R. (eds.) MATERIALS In this work a set of hyperspectral images has been used for the evaluation of the proposed algorithm. Those set comprises three atmospherically corrected airborne HyMAP [1] scenes, two radiometrically corrected spaceborne EO-1 Hyperion [2] scenes and two radiometrically corrected airborne AISA Dual [3] scenes. The HyMAP scenes were acquired over Potsdam (2004), Dresden (2003) and Berlin (2005) in Germany. The Hyperion scenes were acquired over the Haib River Complex crossing Namibia and the Republic of South Africa in 2013 and 2014. The AISA scenes were acquired over the Fichtwald region in Germany in 2010. The HyMAP acquisitions have been utilized as test bed for destriping of artificial stripe degradations, because whisk broom data takes are mostly unaffected by dark current variations induced stripes. Those scenes have been atmospherically corrected using ATCOR [4] before degradation. Then, each band of each scene has been independently and artificially degraded by adding White Gaussian Noise (WGN) that was generated as proposed in [5], whereas different noise degradation levels (0.1 %, 0.5 %, 1 % and 5 %) have been realized resulting in 1536 different striping scenarios (4 noise levels, 3 HyMAP images, 128 bands; figure 2). For each scenario ground-truth was available as original image that then served as evaluation basis. All other hyperspectral scenes from AISA and Hyperion remained unchanged due to their existing stripe degradation as shown in figure 1 and 2.