Details of image radiometric correction
Overall clipping of histogram
Controlling that the available dynamic range is efficiently used with no undesirable saturation helps considerably for the preservation of the original information content needed for the correct manual photointerpretation (CAPI). It is especially helpful to improve readability of the image over dark or bright areas (caused by shadows or highly reflective objects).
The calculation of this metric is straightforward, monitoring whether the rate of pixels belonging to the first and last 5 bins of the luminosity histogram are each less than 0.5% of the total (recommended). In addition, it might be worthwhile to also apply this radiometric ‘clipping’ metric separately for each individual channel (MS and PAN), in the case a VHR bundle ortho-product is used. This is especially valid for the NIR band, since it is a key contributor for the false-color composite. This NIR does not contribute to the luminosity value, and is therefore not assessed by default. Checking the NIR band ‘cut-off’ of the VHR satellite sensors is important as we sometimes observe an over-exposure (saturation) of the grey values for the NIR band over smooth surfaces with high reflectance (partly due to specular effect). More often, such cutting off can also be a result of the rescaling to 8 bit (see images below). By processing with 16-bit files, you maintain all the information necessary to ensure quality in all portions of the histogram.
So, the best practice in imagery use is to maintain its greater bit-depth.
The aim of this parameter is to ensure that the overall lightness of the image is optimised. It controls whether the minimum and maximum extents of the histogram peaks of the different spectral bands are within the recommended +/-15% range from the median value (it is 128 for 8 bit range). Indeed the exact placement and shape of the histogram can vary significantly depending of the image content (important factors are lighting conditions, land cover type, topography, etc.). Therefore, the recommended thresholds cannot be considered as absolute requirements. The aim is not to force the histogram to fit on or converge over a common spot. We could rather apply a shaping function or specific LUT (look-up table) stretch to achieve more balanced visual appearance on screen. Although dedicated to the MS part of the orthoimagery, this metric might be relevant also for the PAN band as well.
The purpose of this parameter is somehow similar to the previous one. It provides a simple quantitative method for assessing whether the tonal neutrality over the orthoimage is met. It can be argued whether or not such numeric verification is needed, since the qualitative visual check often provides good estimates. The problem is that this visual check is made against reference data, which is often not available, or (if present) is not always reliable. In addition, qualitative color balance evaluation is performed only under proper lighting conditions on display systems that are calibrated and have the correct tone scale workstation conditions that cannot be always fulfilled. The proposed metric checks whether the difference between the minimum and maximum digital counts of the RGB pixel values of a given triplet is less than the recommended 2% of the total available bit range (5 bins for an 8 bit image). Even if the metric is very simple to apply, there are certain issues with respect to its representativeness:
The first issue is that the chosen triplet should belong to a spectrally neutral object, such as paved roads or building roofs. The results highly depend on the quality and representativeness of the sample selected. In principle, we expect a neutral object to have the same reflectance intensity along the visible wavelength range, resulting in rather achromatic (grey) colour appearance. However, a reliable visual assessment of such neutrality could be jeopardised by the same negative factors that have impact on the qualitative visual assessment (lighting conditions, calibrated monitor, etc.).
The second issue is that the metric is based on measuring the colour balance on, albeit well-representative, still isolated spots of the image. It would make sense also to check the homogeneity of the object where the triplet is located. The ITT report for NAIP specifies a different metric and threshold for the difference between triplets across an image, within the same object type or cover type (see page 23 of the report). In addition, the report suggests also another metric controlling the colour saturation that is consistently applied on the whole image (see point 4.3 on page 14). Further research is needed to reveal which metric is more relevant in the context of LPIS and CwRS
Finally, we need to be aware that for VHR satellite data the degree of colour preservation the original MS image depends considerably on the rescaling and pan-sharpening methods. Often the appropriate sharpness of the PAN is achieved at the expense of poorer colour recovery or vice-versa.
On picture below, you can see a 16bit image (1), rescaled to 8-bit with Min-Max method (2) and with StDev method (3). Differences for the RGB triplet are within 2% for (2), but not for (3).
The picture below shows the same image pan-sharpened with 3 different algorithms and showing RGB values for the same triplet. We can see the DN values inside the RGB triplet vary depending on the pan-sharpening method.
Noise is defined as non-image related variations in intensity and can have an effect on the interpretation of an image. The Signal to Noise Ratio (SNR) method proposed in the JRC ortho-guidelines are based on the simple ratio between the mean DN value and the standard deviation of the DN values over a given uniform neighbourhood (often calculated on the base of a 3X3 moving window). It should target areas that are prone to be noisier. The Rose criterion (after Albert Rose) states that an SNR of at least 5 is needed to be able to distinguish image features at 100% certainty (see the model.pdf Rose Model). However, it is not yet clear whether this method of calculating the SNR is the optimal in the CAP context. The JRC ortho-guidelines propose that the standard deviation alone as a global statistic over the whole image gives sufficient estimate on the amount of undesirable noise. The recommended upper limit of the standard deviation is 12. Further analysis can be made in selected homogeneous/inhomogeneous areas. Some contractors propose as well the PSNR (Peak signal-to-noise ratio) as an alternative measure. Furthermore, we need also to take into account that the acquisition systems of VHR sensors and the new digital aerial cameras are designed and adjusted in a way to ensure that in the majority of cases, a sufficient ratio between the meaningful input signal and the background noise. In such case, the SNR metric should be probably focused on assessing the quality of the different image compression methods after the ortho production.
Contrast is one of the basic and most essential metrics monitoring the amount of information that an image can provide. High contrast images allow better distinction of land cover feature and facilitate at great extent the work of the CAPI operator. The contrast in the JRC ortho-guidelines is expressed through the coefficient of variation, which is the standard deviation of the DN values as a percentage of the available grey levels. It should preferably be in the range of 10 % - 20%. It should be calculated separately for each band (MS and PAN), in case if VHR bundle ortho-product. The ITT report suggests also an alternative contrast metric (calculated on the luminosity) given in point 4.2 of the report. It derives the number of bins between the minimum and the maximum of the cumulative luminosity histogram function. In principle the modern digital sensors (aerial or satellite) with high radiometric resolution (11bit or more), do not experience problems providing images with sufficient contrast, since the adjustment of their dynamic range during the acquisition and pre-processing ensures that the full histogram extent is obtained for each channel, leaving some margins at the tail for further post-processing. However, the dynamic range of the input imagery depends not only on the light source intensity and target reflectivity, and can become an issue for imagery with noticeable percentage of cloud cover or haze.
Luminance histograms are more accurate than RGB histograms at describing the perceived brightness distribution or "luminosity" within an image. Luminance takes into account the fact that the human eye is more sensitive to green light than red or blue light. The luminance histogram also matches the green histogram more than any other colour.
In order to produce a luminance histogram, each pixel is converted so that it represents a luminosity based on a weighted average of the three colours at that pixel. This weighting assumes that green represents 59% of the perceived luminosity, while the red and blue channels account for just 30% and 11%, respectively. Once all pixels have been converted into luminosity, a luminance histogram is produced by counting how many pixels are at each luminance, identically to how a histogram is produced for a single colour.
An important difference to take away from the above calculation is that while luminance histograms keep track of the location of each colour pixel, RGB histograms discard this information. A RGB histogram produces three independent histograms and then adds them together, irrespective of whether or not each colour came from the same pixel.
When assessing the luminance of VHR satellite data in particular, one should pay attention to the fact that the luminosity histograms generated for the original multispectral image and the resulting pan-sharpened product are different. This is mainly due to the "non-linear" contribution of the panchromatic component in the pan-sharpened product (see the example below) In case no chapter supplied. IMAGE
end of page Radiometric correction
Go back to the Navigation page OTSC