Does my image have glint or adjacency effects?
The short answer is yes.
The long answer is yes, but how much, and is your application sensitive to these effects? If you want to use any of the RedEdge or NIR bands for nearshore or inland waters, e.g. for retrieving turbidity or chlorophyll a, your results will likely be impacted. The severity of the impact will obviously depend on the amount of signal that originates from the air-water interface and/or adjacency. Default masking thresholds in ACOLITE are set to exclude the more severe cases of these effects. The thresholds can be changed by the user, but the retrieval quality may be compromised in cases of high glint and/or adjacency. If you get outputs that are fully masked, these thresholds are the likely culprit (see also viewtopic.php?f=5&t=240 and links therein). You can for example evaluate the required threshold values to remove the masking by looking at the L2R output file in SNAP (https://step.esa.int/main/download/snap-download/), noting the SWIR reflectance of the areas that are masked in the L2W outputs. Some sensors do not have SWIR bands and hence other bands are used for masking, check the sensor specific defaults for details. While changing these thresholds is easy, this is not a general recommendation to do so. Please consider why the thresholds are in use and proceed with caution!
The presence of glint and adjacency in NIR and SWIR bands usually causes failures in typical ocean colour atmospheric correction algorithms (e.g. Gordon and Wang 1994, Antoine and Morel 1999) as the aerosol contribution will be (greatly) overestimated. The DSF algorithm was partly designed to not fail under those conditions, but its outputs will retain most of the glint and adjacency in the estimated surface level signal. Several studies have demonstrated that good results can be obtained in the presence of glint using modelling approaches such as POLYMER (Steinmetz et al. 2011) or neural networks (e.g. C2RCC, Brockmann et al. 2016) if the target water type is well represented by the used model. These methods provide good results in clearer waters, but may perform poorly in turbid waters (e.g. Vanhellemont and Ruddick 2021). In the presence of strong adjacency effects these approaches tend to provide mixed results.
Note that there can also be crosstalk in the SWIR bands especially over dark SWIR targets such as water. This seems to occur especially for Sentinel-2 SWIR2 (B12). This can result in "ghosting" of land features, i.e. a faint outline of land targets or the shoreline shifted in the image. This effect can be visualised by histogram stretching the SWIR2 band to the peak at the low end of the reflectance distribution.
Does my image have glint or adjacency effects?
Re: Does my image have glint or adjacency effects?
How to recognise glint?
Glint generally shifts the spectrum up, with a much higher relative contribution at longer wavelengths. Turbidity retrievals will hence be (significantly) overestimated. Since the glint spectrum is rather flat, retrieving chlorophyll a from Red/RedEdge band ratios will be less severely impacted. The presence of glint depends mainly on the illumination and observation geometry, and the water surface roughness. Glint reflectance tends to increase towards the east of the swath - towards the sun. Small scale variability in the surface roughness of the water can create patterns in the glint, e.g. as a result of local wind variability or differences in water masses.
In this example from Lago Strobel (Santa Cruz province, Patagonia, Argentina) as imaged by Sentinel-2B on 2021-12-23 (scene S2B_MSIL1C_20211223T141729_N0301_R010_T18FYM_20211223T174543), elevated reflectances are found for both pins, with some additional surface variability (likely due to local wind patterns) near Pin 1. The elevated reflectance is especially noticeable as an offset in the NIR and SWIR. The RGB composite uses linear scaling of the surface level reflectance at 665, 559 and 492 nm between 0 and 0.15 to 8 bits per band to form a 24 bit image.
Glint generally shifts the spectrum up, with a much higher relative contribution at longer wavelengths. Turbidity retrievals will hence be (significantly) overestimated. Since the glint spectrum is rather flat, retrieving chlorophyll a from Red/RedEdge band ratios will be less severely impacted. The presence of glint depends mainly on the illumination and observation geometry, and the water surface roughness. Glint reflectance tends to increase towards the east of the swath - towards the sun. Small scale variability in the surface roughness of the water can create patterns in the glint, e.g. as a result of local wind variability or differences in water masses.
In this example from Lago Strobel (Santa Cruz province, Patagonia, Argentina) as imaged by Sentinel-2B on 2021-12-23 (scene S2B_MSIL1C_20211223T141729_N0301_R010_T18FYM_20211223T174543), elevated reflectances are found for both pins, with some additional surface variability (likely due to local wind patterns) near Pin 1. The elevated reflectance is especially noticeable as an offset in the NIR and SWIR. The RGB composite uses linear scaling of the surface level reflectance at 665, 559 and 492 nm between 0 and 0.15 to 8 bits per band to form a 24 bit image.
- GlintLagoStrobel_Glint.jpg (229.25 KiB) Viewed 73678 times
Re: Does my image have glint or adjacency effects?
How to recognise adjacency effects?
Adjacency effects increase the apparent reflectance and depend on atmospheric conditions and the surrounding material. The adjacency effect depends on the diffuse transmittance of the atmosphere and is hence largely determined by the aerosol optical thickness. For the same target with similar surroundings the amount of adjacency effects varies largely with the concentration of aerosols. If your target is surrounded by ice, bare soil, or vegetation, it will experience very different adjacency impacts, due to the differences in spectral shapes of those materials. For vegetation adjacency, the retrievals for visible bands are usually OK, due to relatively low vegetation reflectance in the visible. The RedEdge and NIR bands give problems, due to the RedEdge increase in typical vegetation spectra. In general RedEdge and NIR performance for small inland waters is not good. Due to the spectral shape of vegetation adjacency, RedEdge based chlorophyll a retrievals can be heavily impacted as a result of this vegetation RedEdge. If your spectrum has a RedEdge peak that does not go to zero in the SWIR, this may be impacted by vegetation adjacency. If it has a narrow RedEdge peak (i.e. peaking around 705 nm, and dropping rapidly toward longer wavelengths due to high water absorption), but zero SWIR, you may be safe. For ice and soil adjacency, there are larger impacts across the whole spectrum, and perhaps less in the SWIR for ice adjacency.
In this example from Minnesota (Gull and Pelican lakes, amongst many others) as imaged by Sentinel-2A on 2020-09-21 (scenes S2A_MSIL1C_20200921T171021_N0209_R112_T15TUM_20200921T211952 and S2A_MSIL1C_20200921T171021_N0209_R112_T15TVM_20200921T211952), elevated reflectances are found for both pins, especially notable in the NIR and SWIR. The reflectance increase shows a sharp bump in the NIR, corresponding to a vegetation RedEdge. The RGB composite uses linear scaling of the surface level reflectance at 665, 559 and 492 nm between 0 and 0.15 to 8 bits per band to form a 24 bit image.
An increase of rhot_865 and rhot_1614 is noticeable towards the lake shores (arbitrary stretch), and smaller lakes are almost entirely under strong adjacency influence:
A similar increase is present to a degree in rhot_2202, but note the crosstalk ghosting in the middle of the lakes (arbitrary stretch):
Adjacency effects increase the apparent reflectance and depend on atmospheric conditions and the surrounding material. The adjacency effect depends on the diffuse transmittance of the atmosphere and is hence largely determined by the aerosol optical thickness. For the same target with similar surroundings the amount of adjacency effects varies largely with the concentration of aerosols. If your target is surrounded by ice, bare soil, or vegetation, it will experience very different adjacency impacts, due to the differences in spectral shapes of those materials. For vegetation adjacency, the retrievals for visible bands are usually OK, due to relatively low vegetation reflectance in the visible. The RedEdge and NIR bands give problems, due to the RedEdge increase in typical vegetation spectra. In general RedEdge and NIR performance for small inland waters is not good. Due to the spectral shape of vegetation adjacency, RedEdge based chlorophyll a retrievals can be heavily impacted as a result of this vegetation RedEdge. If your spectrum has a RedEdge peak that does not go to zero in the SWIR, this may be impacted by vegetation adjacency. If it has a narrow RedEdge peak (i.e. peaking around 705 nm, and dropping rapidly toward longer wavelengths due to high water absorption), but zero SWIR, you may be safe. For ice and soil adjacency, there are larger impacts across the whole spectrum, and perhaps less in the SWIR for ice adjacency.
In this example from Minnesota (Gull and Pelican lakes, amongst many others) as imaged by Sentinel-2A on 2020-09-21 (scenes S2A_MSIL1C_20200921T171021_N0209_R112_T15TUM_20200921T211952 and S2A_MSIL1C_20200921T171021_N0209_R112_T15TVM_20200921T211952), elevated reflectances are found for both pins, especially notable in the NIR and SWIR. The reflectance increase shows a sharp bump in the NIR, corresponding to a vegetation RedEdge. The RGB composite uses linear scaling of the surface level reflectance at 665, 559 and 492 nm between 0 and 0.15 to 8 bits per band to form a 24 bit image.
- S2A_MSI_2020_09_21_17_21_01_merged_L2R_adjacency_example.jpg (240.94 KiB) Viewed 73678 times
- S2A_MSI_2020_09_21_17_21_01_merged_L2R_rhot_865.png (100.9 KiB) Viewed 73678 times
- S2A_MSI_2020_09_21_17_21_01_merged_L2R_rhot_1614.png (93.25 KiB) Viewed 73678 times
- S2A_MSI_2020_09_21_17_21_01_merged_L2R_rhot_2202.png (95.69 KiB) Viewed 73678 times
Re: Does my image have glint or adjacency effects?
Can we correct these effects?
If the sensor has SWIR bands, glint effects can be reasonably well corrected by using one of two built in correction methods and by setting dsf_residual_glint_correction=True. These methods use the image itself to derive a per-pixel glint estimate. One method (dsf_residual_glint_correction_method=default) uses the ratio of direct atmospheric transmittances and fresnel reflectance and an observation in the SWIR to estimate a glint spectrum (as in Harmel et al. 2018). The second method (dsf_residual_glint_correction_method=alternative) uses the average SWIR (or NIR if so configured) surface level reflectance and the air-water interface reflectance as modelled using OSOAA and a high wind speed. In the presence of adjacency effects these glint correction methods will interpret the adjacency signal as glint, and will hence overcorrect the image, leading to negative retrievals.
The example from Lago Strobel as corrected with the "default" glint correction:
The example from Lago Strobel as corrected with the "alternative" glint correction:
Note how after the glint correction the spectra are shifted down, and how the reflectances from both pins align much better.
Correction methods for adjacency effects and the combination of glint and adjacency effects are under development. These are difficult problems to solve, and at present results from these methods give mixed performance.
If the sensor has SWIR bands, glint effects can be reasonably well corrected by using one of two built in correction methods and by setting dsf_residual_glint_correction=True. These methods use the image itself to derive a per-pixel glint estimate. One method (dsf_residual_glint_correction_method=default) uses the ratio of direct atmospheric transmittances and fresnel reflectance and an observation in the SWIR to estimate a glint spectrum (as in Harmel et al. 2018). The second method (dsf_residual_glint_correction_method=alternative) uses the average SWIR (or NIR if so configured) surface level reflectance and the air-water interface reflectance as modelled using OSOAA and a high wind speed. In the presence of adjacency effects these glint correction methods will interpret the adjacency signal as glint, and will hence overcorrect the image, leading to negative retrievals.
The example from Lago Strobel as corrected with the "default" glint correction:
- GlintLagoStrobel_GlintCorr.jpg (204.22 KiB) Viewed 73677 times
- GlintLagoStrobel_GlintCorrAlt.jpg (207.05 KiB) Viewed 73677 times
Correction methods for adjacency effects and the combination of glint and adjacency effects are under development. These are difficult problems to solve, and at present results from these methods give mixed performance.
Re: Does my image have glint or adjacency effects?
(post reserved)