^{1}

^{*}

^{1}

^{1}

^{1}

^{2}

^{2}

^{1}

^{1}

^{1}

^{1}

This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution license (http://creativecommons.org/licenses/by/3.0/).

The analysis of hyperspectral images is an important task in Remote Sensing. Foregoing radiometric calibration results in the assignment of incident electromagnetic radiation to digital numbers and reduces the striping caused by slightly different responses of the pixel detectors. However, due to uncertainties in the calibration some striping remains. This publication presents a new reduction framework that efficiently reduces linear and nonlinear miscalibrations by an image-driven, radiometric recalibration and rescaling. The proposed framework—Reduction Of Miscalibration Effects (ROME)—considering spectral and spatial probability distributions, is constrained by specific minimisation and maximisation principles and incorporates image processing techniques such as Minkowski metrics and convolution. To objectively evaluate the performance of the new approach, the technique was applied to a variety of commonly used image examples and to one simulated and miscalibrated EnMAP (Environmental Mapping and Analysis Program) scene. Other examples consist of miscalibrated AISA/Eagle VNIR (Visible and Near Infrared) and Hawk SWIR (Short Wave Infrared) scenes of rural areas of the region Fichtwald in Germany and Hyperion scenes of the Jalal-Abad district in Southern Kyrgyzstan. Recovery rates of approximately 97% for linear and approximately 94% for nonlinear miscalibrated data were achieved, clearly demonstrating the benefits of the new approach and its potential for broad applicability to miscalibrated pushbroom sensor data.

The potential of imaging spectroscopy to provide more and better information about the Earth than do multispectral instruments is currently accompanied by an intensified development and availability of new hyperspectral airborne and spaceborne sensors. The new generation of hyperspectral sensors utilise the pushbroom technology, enabling an integration time per detector element. Therefore, these sensors obtain a better Signal-to-Noise Ratio (SNR) compared to whiskbroom scanners. However, the use of detector arrays in the sensor design requires a more precise radiometric calibration. Even small variations will cause striping effects in the image data that aggravate subsequent analyses such as classification and segmentation [

The rescaling aims to adjust the radiometric scale by considering areas of lowest reduction. This is necessary since uncertainties remain in the estimation of parameters (e.g., detector resolution in the linear slope reduction) and in the incorporation of miscalibrated reference areas (e.g., potential miscalibration of the first image column as reference for the offset reduction).

Additionally, it will be shown how potential trends or frequency undershoots caused by corrections themselves or by low SNR can be suppressed. In addition, an adapted data dimensionality reduction is proposed, which desensitises striping reduction approaches in the presence of edges and increases computational speed. Here, Minkowski metrics, gradient operators and edge extraction algorithms are combined to exclude discontinuities such as edges and impulse noise from further analyses if they do not dominate the image content [

Four grey valued images from the image database of the Signal and Image Processing Institute (SIPI) of the University of California [

To simulate different types of linear miscalibration, each of the four grey valued images from the SIPI image database (

For every specific noise degradation level and type (slope, offset) the Gaussian white noise was randomly generated [

Each set was applied to each grey valued image and consists of 5 variations representing the different noise types slope and offset, slope only, offset only, slope and a priori knowledge and offset and a priori knowledge. In the result each grey valued image was 400 times differently degraded.

The rescaled multiplicative noise ranged in maxima from 1 to 1789 (mult1), whereas the minimum was fixed to 0.0001 (mult2) to enable an impact of the multiplicative degradation. The rescaled additive noise ranged in maxima from 5.59 to 10,000 (off1) and in minima from −5.59 to −10,000 (off2). The rescaling was based on exponential functions to simulate small and large degradations. To detect potential scaling effects, the scaling was varied four times within the 80 different noise levels:

Case 1: multiplicative from mult2 to mult1 and additive from mult2 to reversed off1.

Case 2: multiplicative from mult2 to mult1 and additive from mult2 to off1.

Case 3: multiplicative from mult2 to mult1 and additive from reversed off2 to reversed off1.

Case 4: multiplicative from mult2 to mult1 and additive from off2 to off1.

A set of six specific miscalibrated hyperspectral images were additionally destriped to test the proposed approach on images that were acquired either from aircraft or from satellite and degraded by either linear or nonlinear miscalibrations. For this purpose, three hyperspectral AISA DUAL scenes [

The three AISA DUAL scenes were acquired on September 23rd 2010 between 1 p.m. and 3 p.m. for the ‘Fichtwald’ study region in Eastern Germany (

The AISA DUAL system consists of two separate pushbroom sensors, AISA Eagle (400–970 nm) and AISA Hawk (970–2450 nm), which are mounted on a stabilised aircraft platform [

The two hyperspectral Hyperion image scenes [

Due to SNR considerations, only 198 bands are routinely processed for generating level 1 images [

The approach was tested further by using data generated to be like those from future EnMAP sensor [

This simulator is used for the optimisation of instrument specifications and the development and validation of data processing and calibration algorithms. An example of a simulated EnMAP image is depicted in

Different physical detector characteristics of a pushbroom sensor produce image stripes in acquired raw data. These stripes are then corrected by radiometric in-flight, vicarious [_{e}^{−} is the number of collected electrons, τ is the optical transmission, λ is the centre wavelength, is the quantum efficiency and F is the filter efficiency. However, the detector signal must be related to a recordable and transmittable digital number (DN), which may be given by the following equation:
_{max} is the radiometric resolution, FWC is the Full Well Capacity that defines detector saturation, and DN_{0} is the dark current. Subsequent laboratory measurements are then used to estimate transformation parameters, either for the transformation of at-sensor radiance L to digital number DN considered to be radiometric calibration or,

Calibration measurements are performed by sensing known physical targets and by a subsequent evaluation of the sensing results. The association is often realised by a polynomial least squares fit, which minimises the differences between modelled and measured at-sensor radiance [_{targets} denotes the number of calibration targets, c_{0} is the offset contrary to the dark current, and M is the polynomial degree. In many cases, the assignment is performed linearly, that is, M becomes 1. It is not possible or practical to calibrate a sensor for each image acquisition, which results in relying on calibration parameters that may no longer be up-to-date. This can lead to vertical stripes requiring a new calibration based solely on image data. Additionally, it is necessary to assess the stripe type to perform the right reduction—multiplicative or additive—linear or nonlinear. This can be achieved by an inspection of the outputs of different reduction approaches, either manually or automatically, which may be more reliable if the differences are small.

Striping reduction can be divided into two types—an additive c_{0} and a multiplicative c_{1..M} reduction. Because how an image is degraded is unknown, all c_{0}, c_{1} and c_{2..M} reductions should be performed and evaluated. The evaluation can be based on the assessment of the

Discontinuities such as edges and impulse noise have a specific impact on the assessment of stripes, which often leads to an exclusion of edges. This generalisation is useful if the spatial contribution of edges is low compared to homogenous regions. Then, the incorporation of edges, which are not strictly across track or along track, can lead to uncertainties in the estimation of striping magnitudes. However, if the spatial contribution of edges is relatively high, it may be not advisable to exclude them.

Exclusion would increase the uncertainties of any stripe assessment if the amount of remaining data is too low. To enable a robust decision, whether edges are incorporated or not, it is necessary to mask them beforehand.

A binary edge mask can be obtained for Remote Sensing Images (RSI) with wavelength-independent striping by applying the Hyperspectral Edge Detection Algorithm (HEDA) as proposed in [

Gradient estimation by convolving image with the derivative of a two-dimensional Gaussian.

Non-Maximum-Suppression of all edge pixels (edgels) which absolute gradient magnitude is lower than the magnitude of adjacent non edge pixels in perpendicular edge direction.

Hysteresis for all remaining edgels by tracing and thresholding edgels for given criteria.

To avoid uncertainties in choosing steering parameters of the Canny or HEDA algorithm and to minimise overestimations of across-track edges of heterogeneous regions, morphological dilations with small rectangular discs as structuring elements are performed [_{w} is a vector of width w valued 1, ⊗ is the dyadic product, ⊕ is morphological dilation,
_{striped} in striping (vertical) direction r and HEDA_{striped} is the HEDA adaption as given previously. The multiplicative application of the inverse edge map on L_{striped} then gives a HSI where high contrast edges are zeroed. This can be performed straightforwardly for each band as shown by the following equation:

To avoid uncertainties in the application of succeeding destriping approaches, masked or missing data as well as neighbour pixels are excluded. The workflow for masking of discontinuities of a hyperspectral image is then given by:

Computation of binary edge maps by the Hyperspectral Edge Detection Algorithm (HEDA) in striping direction (

Morphological dilation to suppress edge related adjacency effects (e.g., PSF related blooming of edge spectra into homogeneous regions) as given by

Binary filtering of the striped band by applying reciprocal binary edge map (

If a scene constant, band and detector-element-dependent slope c_{1} is assumed, then c_{1} contributes to each element (pixel) within a column of one band the same multiplicative fraction (to avoid confusion in this work, the term ‘gain’ corresponds to the maximisation of the radiometric resolution [_{1} slope for each column and band can then be performed in multiple steps whereby a least squares polynomial fit is not required. In relation to _{1} slope and to the difference of detected radiation equivalent DNs, because the offset c_{0} is constant per column and band and is eliminated in such a relation ( c1*DN1+ c0 – c1*DN2−c0 = c1*(DN1−DN2) ).

This basically reduces the mathematical complexity in the linear slope reduction case down to the retrieval and evaluation of the distribution of differences. Therefore, an elaborated polynomial fit is not necessary in the linear case. Because most miscalibrations can be corrected efficiently by linear reductions, the slope reductions are divided into a linear and a nonlinear case, whereby a potential multi-step approach for linear reduction is described in the following.

In the first step, the grey values or radiances for each band and column are sorted in ascending order. In the second step, spectrally unique values _{s} are extracted for each band by:
_{nr} denotes a column index vector ranging from 1 to nr of length nr, s denotes the sort index and
_{s} shifted backwards in place by one in row direction.

In a third step, the differences for each column of unique column values are used to detect the potential c_{1} slope of this column, which can be defined as:
_{1}(diff) always gives the smallest difference. This smallest difference is equivalent to c_{1} times the smallest difference of the unique values (SDUV) for each band and column of a not-striped representation of the striped image (perfectly calibrated). SDUV can be also interpreted as detector resolution. The resulting equation for the column-based estimation of the c_{1} slope is then given to:

If SDUV is one (e.g., often for grey valued images) then both sides of relation (8) are equivalent or else further processing necessary in which SDUV must be estimated.

Such an assessment can base on the median of column differences. This then gives the final relation for the estimation of c_{1} in short notation to:

In the fifth step, the application of the c_{1} slope reduction is verified. At first, the column c_{1} is compared to one to avoid unnecessary reductions. In this step, c_{1} is compared to c_{1} of the right adjacent column. This is performed by evaluating the difference in the related histograms for both columns. If the number of histogram bins and the positions of the maxima are equal for both columns, then c_{1} reduction should be not applied for that column, because it can be assumed that different c_{1} cause ‘stretches’ of the histogram and a shift between their maxima.

If these exclusion conditions are not fulfilled, then c_{1} is applied for that column by division. If there are also offset stripes, then these offsets are reduced concurrently to c_{0}/c_{1}, which can be minimised by a succeeding offset reduction. Subsequently, a rescaling of the data is required if offset stripes are present. Exclusion conditions and c_{1} reductions should be recorded concurrently to enable restoration of the original data, if cSNR indicates a necessary revoking of c_{1} reduction. Due to uncertainties in the assessment of SDUV, a linear radiometric rescaling should be performed after linear c_{1} reduction which is proposed in section 3.6. The respective workflow for the linear slope reduction per band is then given by:

Sorting of column radiances.

Extraction of unique column radiances (

Calculation of differences of unique column differences (

Estimation of slope per column as ratio of the smallest difference of unique values of a striped band (

Verification of the slope reduction necessity by evaluation of the shapes of the histograms of adjacent columns.

In the following, a multi-step approach for the reduction of offset miscalibrations is proposed that is conditioned by the assumptions that a radiometric offset is detector-element dependent and varies from scene to scene but not within a scene. This reduction approach consists of three steps that are applied consecutively. In the first step, two adjacent columns are considered. If these two columns cover a small, homogeneous area with assumed equal surface cover type, viewing geometry and second order effects [_{c0} contains the offset difference which is given in the following to:
_{striped,flat} denotes the striped hyperspectral image without edges (compare

From this, it follows that remote sensing scenes may be especially convenient due to their along-track size. It also follows from this that a cluster agglomerating the majority of differences within the difference vector may most likely contain the offset reduction coefficient which is the basic assumption of this offset reduction approach. Thus, the distribution of the difference vector is examined in a normalised histogram in the second step. At first, the histogram is sorted in descending order according to its frequency as given in the following:

Then, the number of the bin or the frequency category of the sorted histogram contributes to the rank of the probability of containing the offset reduction coefficient. In the third step, the first N bins of the differences histogram are considered, whereas N is user given and should be greater than or equal to 1. A representative offset reduction coefficient for each bin can then be obtained by computing the median value for each bin as given as follows:
_{i} denotes the sorted frequency category i.

This concurrently reduces uncertainties caused by pre-processing such as the discontinuity masking. Subsequently, these bin reduction coefficients are weighted according to their frequencies. The median of the weighted bin correction factors is then supposed to be the offset correction coefficient of this column and is immediately subtracted. Coevally, the offset reduction coefficient is stored to rebuild the original data if succeeding cSNR estimation indicates a different stripe type. To reduce uncertainties in the determination of the offset deviation of the first column, a subsequent linear radiometric rescaling for each band is necessary, as is proposed in section 3.6. The workflow of the offset reduction per band can be summarised by:

Calculation of the differences of adjacent columns (

Descending sorting of the estimated probability distributions of the column differences (

Calculation of the average column difference per frequency category (

Weighting of averaged column differences by their normalised frequencies.

Estimation of the column offset as average of all weighted average column differences in low SNR scenarios or as first weighted average column difference of a descending frequency sorted histogram of column differences in high SNR scenarious.

Nonlinear miscalibrations must incorporate correcting coefficients of higher degrees. It follows from this that additional repetitive information is crucial for the estimation of coefficients. Additionally, the distribution and scaling of the respective radiance domain is not known, which aggravates a polynomial radiance assignment. In the following, a multi-step band-wise approach is proposed, reducing both the influence of these limitations and nonlinear striping. It consists of a linear c_{1} and a nonlinear c_{2..M} reduction as well as a c_{0} reduction.

At first, the minimum and the maximum radiance per band as well as all unique column radiances per band are determined as described in Section 3.3. Secondly, the distribution of all column frequencies is considered. For this, a histogram can be used that is limited to the minimum and maximum radiance in this band and is binned with a frequency interval equal to the difference between maximum and minimum divided by the number of unique radiances of this band. This frequency interval then gives a domain for all unique column radiances for which it is assumed that dark current and saturation are similar for all detectors. Finally, unique column radiances are considered based on their frequencies in the previously defined histogram. The bin numbers or frequency interval numbers themselves then gives the domain per column, which could be considered to be quasi DN, and the unique column radiances can be fitted against their domain. For this purpose broadly used nonlinear least squares fits may be used. To avoid uncertainties in the determination of column offsets, the subsequent column-related reduction of nonlinearities by subtraction should be only applied on all column radiances with polynomial column coefficients greater than one. After this, a linear slope and offset correction should be applied, as previously described. Additionally, it is necessary to recover the radiance level and the radiance scaling of the entire band as described in Section 3.6. The respective workflow for the nonlinear reduction per band is given by:

Detection of minimum and maximum per band and extraction of unique column radiances (

Definition of a histogram domain (x-axis) in respect to band maximum and minimum.

Extraction of frequency categories (x-vectors) per unique column radiances in respect to the previously defined histogram domain.

Least squares polynomial fit of column frequency categories (x-axis) and unique column radiances per column (y-axis).

Reduction of column nonlinearities by subtracting estimated column polynom c_{2..M}.

Linear slope and offset reduction as given in Sections 3.3 and 3.4.

In order to reduce various effects of foregoing reductions a spectral rescaling is necessary. These effects can be caused by:

c_{0} offset reduction comprising offset relations.

linear c_{1} reduction potentially biased by differences between real and assessed SDUV.

nonlinear c_{2..M} reduction assuming similarities of band minima and dark currents as well as band maxima and saturations.

They can be minimised by rescaling the destriped radiance spectra to minimally striped areas. For this purpose a multistep approach was developed that detects and evaluates areas of lowest offset striping quantity. Based on these areas the whole band is rescaled by considering the spectral range before and after destriping.

At first, a striping quality indicator (SQI) is defined that combines both the level of stripes and their variation within a window. SQI can be defined as a vector of products of standard deviations of absolute reductions times the median of absolute reductions within a window of a pre-defined size (3 columns in size in minimum). Secondly, minima of SQI vectors are identified to detect minimal striped areas.

The position of the minimum is then indicating lowest reduction. The middle of the window is used as positional index for the column used as reference. To avoid rescaling of destriped images in reference to columns within this window, which are significantly miscalibrated, the reduction quantities should be considered. This can be performed by an evaluation of the ratio between the mean of first and last reduction coefficient within the minimum window and the reduction coefficient at the positional index. If all criteria are fulfilled, that is, a striping reduction was applied, SQI gives a positional index of minimal striped area, where striping is significant lower in the window, centred at the positional index, compared to other areas. Then, the last step—the scaling—is applied for each band at the positional index or column by:
_{old} as local minima and max_{old} as local maxima in the destriped image window and min_{ref} as local minima and max_{ref} as local maxima in the striped image window similar to [

Computation of SQI within a moving window of a pre-defined size.

Definition of positional index by detecting the minimum within the SQI vector.

Rescaling of the whole band by applying

In specific cases, brightness gradients of the destriped image can additionally disturb succeeding analyses. These trends may be caused by offset reduction undershoots or by material dependent effects due to varying illumination and acquisition geometry. In the following, a multi-step, band-related approach is proposed, which aims at the reduction of these effects. First, the median of each column is calculated, which is robust in the presence of outliers. This median vector _{med} can be then boxcar convolved to get a smoothed vector representation _{norm} denotes the mean normalised smoothed median vector, _{med} denotes the median vector, * denotes convolution and w the boxcar or window size. The workflow for detrending is then given by:

Calculation of the median vector consisting of the median values for each column.

Smoothing of the median vector to remove outliers.

Mean normalisation of the smoothed median vector to distinguish trends.

Detrending the rescaled, destriped band by applying

Some of the sub steps of the previously proposed approach are limited to one spectral dimension. In consequence, it is recommended to destripe multi-dimensional data for each band as shown in

The global Peak-Signal-to-Noise-Ratio (PSNR) [

Procedures to reduce radiometric miscalibration were automated to be repeatable.

The four grey valued images were selected to cover a broad range of spatial and spectral contributions. The ‘Lenna’ image provides a homogenous grey value distribution but contains also a long ‘natural’ vertical stripe in the left part of the image. The ‘Mandrill’ image is imbalanced due to spatial distribution of gradients and spectrally homogeneous regions. The ‘Aerial’ image has a homogenous grey value distribution, but lacks grey value variance. The ‘Sailboat on lake’ image can be characterised as having a balance of gradients and homogeneous regions as well as a homogeneous grey value distribution. Each grey valued image was differently striped 400 times and then destriped (see examples of

All three image quality indicators—PSNR, Entropy and MSSIM [

Approximately 93% (100–7.3%) of the original information content was recovered (_{1} slope reduction outperformed offset reduction, whereby a priori knowledge had the highest impact on slope-related destriping. A visual inspection of all destriping results indicated that approximately 7% of them exhibited barely any detectable stripes that were potentially related to the information content quantity that could not be recovered. This supports the assumption that the combination of image quality indicators allows for the more efficient evaluation of image processing outputs [_{0}-reduction approach (

According to the results presented in

On average, very high and robust image recoveries were achieved. This is exemplarily demonstrated for weak striping (e.g.,

The destriping results of the hyperspectral scenes were evaluated like the grey valued images. Nevertheless, two cases must be differentiated. Ground truth was available only for the EnMAP scene, and this scene was nonlinearly, artificially striped. The AISA and the Hyperion scenes were linearly destriped in association to the cSNR-related decision system.

The averages of image quality indicators are shown in

All bands of inspected AISA DUAL scenes showed miscalibrations (e.g.,

As previously assumed, miscalibrations may vary only slowly over time. This indicates that corrections are highly correlated and that they show similar stripe patterns concurrently, which could be addressed by considering the gradients of removed stripes.

To enable a scene- and sensor-independent estimation of the stripe metrics of striped AISA and Hyperion data, the across-track gradients of the detected offset stripes were evaluated. At first, the gradients of the difference of the striped and the destriped images were computed for each band. Then, a random row of each gradient scene was selected. After this, the correlation coefficients between each scene and band were computed. Additionally, least squares regression coefficients for the correlation vector of each scene pair were estimated to detect potential trends. The average stripe correlations as well as the regression coefficients clearly indicate for the AISA DUAL and Hyperion scenes that the reductions are stable over time (

Missing values in the Hyperion scenes were interpolated by Piecewise Cubic Hermite Interpolation Polynomials in across- and along-track directions [

In contrast to the AISA and Hyperion scenes, the EnMAP scene contains nonlinear, artificial effects and hence stripes appear (e.g., in

According to the results shown in

The striped representation of the EnMAP ground truth scene had an image quality deviation to ground truth of approximately 22%, that is, the overall image quality improvement between the striped and the striping reduced scene was approximately 21%. However, a subjective visual examination of the recalibrated EnMAP scenes has revealed that not all stripes were removed, corresponding to a detailed comparison between ground truth and destriped transects (

The proposed ROME framework is potentially affected by low SNR scenarios (SNR less than 20) where the offset reduction can cause frequency undershoots that are minimised by the proposed trend correction. However, sensitivity investigations have revealed that reduction quality is only weakly correlated to SNR.

Another aspect is the influence of natural, column-parallel, homogeneous regions covering the whole image extent on the offset reduction. Although the occurrence of such structures is not very likely for the majority of image data, they could be automatically detected and separately processed by repeated rescaling of these regions in order to avoid related artefacts.

The developed ROME framework for the reduction of linear and nonlinear miscalibration effects consists of two main parts—the spatial reduction of striping and the spectral rescaling of the image data. In order to reduce stripes and to preserve the spectral characteristics of observed surfaces both parts have to be executed consecutively.

The impact of these steps is proportional to the impact of the miscalibration on post-processing. Further processing steps, such as atmospheric correction and classification showed that uncorrected striped or miscalibrated data strongly affect subsequent analyses and result among others in striped columnar water vapour and aerosol optical thickness estimations and wrong segmentations. Hence, destriping is strongly recommended if data contain visually perceptible stripes after calibration.

The developed approach was widely tested and evaluated by different methodologies. In this process high calibration recovery rates of approximately 97% for linear miscalibration and approximately 94% for nonlinear miscalibration have been achieved. It was shown that linear and nonlinear miscalibration-related striping can be efficiently suppressed without a significant loss of radiometric scaling and variance information or gradient magnitudes information. Furthermore, destriping did not lead to artefacts in the resulting image spectra which is exemplary demonstrated by the profiles shown in

Concurrently, the successful destriping of images of different origin, stripe type and magnitude demonstrates the broad applicability, robustness and the high performance of the developed approach.

Additionally, the full automation, the reduced mathematical complexity of the proposed method, as well as the insensitivity to

This work was supported by the MoLaWa project funded by the German Federal Ministry of Economics and Technology (BMWi) in the framework of the program Zentrales Innovationsprogramm Mittelstand (ZIM).

Grey-scaled representations [

Study regions: (

False coloured RGB image [band 12–479 nm (blue), 65–801 nm (green) and 213–2201 nm (red)] of the ‘Makhtesh Ramon’ study region as a simulated EnMAP scene.

The destriping processing chain of the ‘ROM ’ framework.

Exemplary striped grey scaled images (left) and destriping results (right) for slope c_{1} and offset c_{0} (a,d,g,j), slope c_{1} (b,e,h,k) and offset c_{0} reductions (c,f,i,l); m=0.0001.

Exemplary grey scaled, striped images of sections in the middle of an AISA DUAL scene (VNIR-band 65–541 nm

Exemplary grey scaled, striped images of sections in the middle of a Hyperion scene (VNIR-band 11–457 nm

Exemplary grey scaled, striped images of sections in the middle of a EnMAP scene (VNIR-band 53–713 nm (

Average deviation of the 1600 destriped images from ground truth [%].

^{*} |
||||||||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|

^{1} |
^{2} |
^{3} |
^{4} |
^{1} |
^{2} |
^{3} |
^{4} |
^{1} |
^{2} |
^{3} |
^{4} | |||

| ||||||||||||||

16.4 | 13.0 | 1.6 | 4.2 | 7.5 | 29.6 | 2.8 | 2.7 | 9.6 | 9.0 | 0.7 | 2.8 | |||

| ||||||||||||||

o | 15.2 | 13.0 | 4.7 | 10.6 | 7.0 | 29.6 | 8.4 | 5.1 | 9.0 | 9.0 | 2.2 | 8.6 | ||

x | 16.8 | 13.0 | 5.1 | 14.5 | 7.7 | 29.5 | 9.2 | 7.1 | 9.7 | 9.0 | 1.9 | 8.1 | ||

| ||||||||||||||

o | 16.8 | 13.0 | 0.0 | 0.0 | 7.7 | 29.6 | 0.0 | 0.0 | 9.7 | 9.0 | 0.0 | 0.0 | ||

x | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | ||

| ||||||||||||||

Flag denotes by ‘o’ no and by ‘x’ a-priori knowledge of the stripe type;

Im 1 is the ‘Lenna’ image;

Im 2 is the ‘Mandrill’ image;

Im 3 is the ‘Aerial’ image;

Im 4 is the ‘Sailboat on lake’ image.

Average stripe magnitude impact on this approach for all destriped images [%].

Im 1^{1} |
Im 2^{2} |
Im 3^{3} |
Im 4^{4} |
Im 1^{1} |
Im 2^{2} |
Im 3^{3} |
Im 4^{4} |
Im 1^{1} |
Im 2^{2} |
Im 3^{3} |
Im 4^{4} | |||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|

| ||||||||||||||

0.8 | 0.0 | 2.1 | 6.2 | 0.8 | 0.0 | 2.1 | 6.2 | 0.4 | 0.0 | 3.8 | 4.7 | |||

| ||||||||||||||

o | 1.7 | 0.0 | 3.1 | 7.7 | 1.7 | 0.0 | 3.1 | 7.7 | 0.9 | 0.0 | 5.9 | 5.7 | ||

x | 0.4 | 0.0 | 3.4 | 9.8 | 0.4 | 0.0 | 3.4 | 9.8 | 0.2 | 0.0 | 6.5 | 7.2 | ||

| ||||||||||||||

o | 0.4 | 0.0 | 1.6 | 4.0 | 0.4 | 0.0 | 1.6 | 4.0 | 0.2 | 0.0 | 2.9 | 2.8 | ||

x | 14.1 | 11.5 | 1.6 | 4.0 | 14.1 | 11.5 | 1.6 | 4.0 | 8.2 | 42.0 | 2.9 | 2.8 | ||

| ||||||||||||||

Flag denotes by ‘o’ no and by ‘x’ a priori knowledge of the stripe type;

Im 1 is the ‘Lenna’ image;

Im 2 is the ‘Mandrill’ image;

Im 3 is the ‘Aerial’ image;

Im 4 is the ‘Sailboat on lake’ image.

Image quality indices for destriped hyperspectral scenes as rate of change [%].

| |||||
---|---|---|---|---|---|

Scene 1 | −0.8 | 5.0 | 9.2 | ||

Scene 2 | −2.1 | 10.1 | 8.3 | ||

Scene 3 | −1.6 | 8.6 | 7.7 | ||

| |||||

| |||||

Scene 1 | 1.2 | 3.1 | 6.0 | ||

Scene 2 | 1.5 | 5.1 | 6.8 | ||

| |||||

| |||||

Scene 1 |

Inter-scene striping relations.

^{2} | ||||
---|---|---|---|---|

| ||||

Scene 1 to Scene 2 | −0.0013 | 0.84 | ||

Scene 1 to Scene 3 | −0.0021 | 0.93 | ||

Scene 2 to Scene 3 | −0.0012 | 0.85 | ||

| ||||

Scene 1 to Scene 2 | −0.0001 | 0.92 |