DESTRIPING OF MISCALIBRATED AISA IMAGES C. Rogaß a,* , D. Spengler a , M. Bochow a , K. Segl a , A. Lausch b and H. Kaufmann a a Section 1.4 Remote Sensing, Helmholtz Centre Potsdam – GFZ German Research Centre for Geosciences, Telegrafenberg, 14473, Potsdam, Germany –(christian.rogass, daniel.spengler, mathias.bochow, karl.segl, hermann.kaufmann)@gfz-potsdam.de b Department of Landscape Ecology, Helmholtz Centre for Environmental Research – UFZ, Permoserstr.15, 04318, Leipzig, Germany - angela.lausch@ufz.de Keywords: radiometric; miscalibration; stripes; hyperspectral; AISA; MoLaWa Abstract: The analysis of hyperspectral images belongs to the main tasks in Remote Sensing. The foregoing linear radiometric correction of registered digital numbers basically assigns the spectral and spatial dependent response of a hyperspectral pushbroom sensor to a physical meaning - radiance. Slopes and offsets of the correction are often determined in laboratory and in-flight calibrations, but may vary over time. This results in striping artefacts which aggravates succeeding processing steps such as atmospheric correction, classification and segmentation. In this work, a new approach is presented, that automatically removes these stripes calculating improved calibration factors without any prior knowledge or user interaction. The algorithm is based on the assessment of spectral and spatial probability distributions and is constrained by specific minimisation principles. Morphological and spatial filtering techniques and additionally a Signal-to-Noise-Ratio related decision tree are implemented to reduce computational effort and to stabilise the solution depending on local spatial entropy. To objectively evaluate the performance of the new approach, the technique was applied to broadly used image processing examples that has been artificially and randomly degraded by sets of multiplicative and additive noise of different distributions as well as miscalibrated AISA DUAL (VNIR and SWIR) scenes. The results clearly show the benefits of the new approach and, concurrently, provide correction facilities for other miscalibrated pushbroom sensor data. 1. Introduction Hyperspectral pushbroom sensor use detector arrays for image acquisitions. Different properties of the detectors demand a precise radiometric calibration. The radiometric calibration assigns known incident at-sensor radiance to measured digital number (DN). The association is often realised by a linear least squares regression. The estimated regression coefficients (slope and offset) are the used in the reverse process – the radiometric scaling - to assign measured DN to unknown incident at-sensor radiance. Uncertainties in the estimation of regression coefficients or variations in the responses of detectors in comparison to last calibration may lead to miscalibrations that are visually perceptible as image stripes. However, miscalibrations aggravate subsequent analyses and have to be reduced (Datt et al., 2003). Miscalibration can be divided into two basic types – additive (offset) and multiplicative (slope) degradation. Offsets are used to incorporate detector-dependent dark current. In contrast, slopes are used to directly assign radiance to DN. The aims of any striping reduction should be stripe suppression and preservation of the physical fingerprint of imaged surface material. In this work a multistep approach is proposed which significantly reduces stripes and preserves the spectral characteristics of imaged surface cover materials. It consists of a linear slope reduction and an offset reduction as well as of specific post-processing steps that are consecutively executed and interim evaluated by the evolution of the Signal-to-Noise-Ratio (SNR). If a previous reduction step has lowered the SNR, then this step will be revoked. This is useful to avoid overcorrections. Both reductions are performed per band. The slope reduction considers a single detector element and the offset reduction relates to adjacent detector elements. Spatial and spectral probability distributions are incorporated which is supported by striping related redundancies. After the reductions of stripes a spectral rescaling is performed that aims to adjust the spectral level of a band by considering areas of lowest reduction. Remaining trends or reduction related frequency undershoots are suppressed in a next, detrending related step. To desensitise proposed reductions in presence of edges, an edge filtering approach was developed that excludes edges before succeeding reductions if they don’t dominate or represent the striped image. For this, Minkowski metrics, gradient operators and edge extraction algorithms were combined (Canny, 1986; Haralick et al., 1987; Rogass et al., 2009). To study the impacts of different linear miscalibrations on the performance of the proposed method, a specific set of grey valued images was randomly striped by linearly varying the slope and/or offset. In addition, a set of 3 hyperspectral, miscalibrated AISA DUAL (SPECIM, 2011) scenes were processed.