Algorithm Specification

Neural network classification

The IdePix classification algorithm for Sentinel-3 OLCI is based on a neural network approach. A common neural net is used for both land and water pixels. As input for the neural net, the square roots of the OLCI TOA reflectances (obtained from an internal radiance-to-reflectance conversion) in all 21 bands are used. It is described in detail in the ATBD [1].

As output, the neural net finally provides per pixel one of the properties 'cloud sure', 'cloud ambiguous', 'cloud' (which means sure OR ambiguous) , or 'snow/ice'.
Over land the 'snow/ice' flag is refined by spatial and spectral tests to provide a better separation between cloud and snow.

Additional properties

The following additional pixel properties are provided from the classification:

'cloud_buffer': For pixels identified as cloudy, a cloud buffer of specified width can be set by the user to finally obtain a more conservative cloud mask. This is done in a post-processing step after the cloud classification has been applied on the full image. The cloud buffer algorithm works on pixel windows of size (2N+1) x (2N+1) with N = cloud buffer width. Note that the cloud buffer is only applied on cloud-free pixels, i.e. cloudy pixels are not flagged as cloud buffer of cloudy neighbour pixels. The cloud buffer procedure is described in a bit more detail in the CCI Land Cover ATBD [2].
'bright': the pixel is a bright pixel. The information is retained from the quality flags of the L1b product.
'land': the pixel is located over land
'coastline': the pixel is a coastline pixel

By default, the 'land' and 'coastline' pixels are identified from the land flag included in the quality flags which come with the L1b product. Optionally, the SRTM (Shuttle Radar Topography Mission) land/water mask [3] can be applied instead. The latter is a fractional mask:

water fraction = 0 : 'land'
water fraction > 0 and water fraction < 100: 'coastline'
water fraction = 100: 'water'

Two shadow flags are provided with the classification:

'cloud_shadow': the cloud top height is estimated from the O2 absorption bands with a dedicated neural network. Starting from each cloud free pixel, the pixels along the (projected) path towards the sun are checked. If they are flagged as 'cloud' and the cloud top height exceeds the elevation of the search path above ground, this cloud is likely to cast a shadow at the start point. The minimum elevation is fixed at 300m above ground. If a cloudy pixel is found within the search path, the 'cloud_shadow' flag is raised at the start pixel and the search is aborted. Otherwise, the algorithm carries on until the maximum length of the search path is reached, which is defined by a theoretical maximum cloud top height at current latitude. The cloud shadow algorithm is applied as part of the postprocessing of the classification.
'mountain_shadow': a hillshade algorithm [4] is implemented, which derives slopes from elevation data and uses a geometric test with slope, aspect and illumination and satellite observation conditions against a threshold to find pixel in the (core) shadow of mountainous terrain. The threshold is set to 0.9 by default, and can be adjusted by the user between 1 (larger terrain shadow extent) and 0 (no terrain shadow).

Final classification flags

In summary, from the algorithm steps outlined above, the following final classification flags are provided for Sentinel-3 OLCI:

'INVALID' (input invalid or no classification possible)
'CLOUD' (which means either CLOUD_AMBIGUOUS or CLOUD_SURE)
'CLOUD_AMBIGUOUS'
'CLOUD_SURE'
'CLOUD_BUFFER'
'CLOUD_SHADOW'
'SNOW_ICE'
'BRIGHT'
'COASTLINE'
'LAND'
'MOUNTAIN_SHADOW'

Known issues

The mountain shadow algorithm can show an artefact at the nadir line, where viewing angles are not exact enough. The following pixel property is currently NOT provided for Sentinel-3 OLCI:

'WHITE'

References

[1] Algorithm Theoretical Basis Document for IdePix Sentinel-3 OLCI
IdePix for Sentinel-3 OLCI - Algorithm Theoretical Basis Document

[2] CCI Land Cover - Algorithm Theoretical Basis Document: Pre-Processing
Chapter 4.4.11 Cloud shadow and cloud edge detection
ESA Landcover CCI page - ATBD: Pre-Processing

[3] Farr, T. G., et al., The Shuttle Radar Topography Mission, Rev. Geophys., 45, RG2004, doi:10.1029/2005RG000183. (2007)
The Shuttle Radar Topography Mission

[4] Hillshade explained
'How Hillshade works' in the ArcGIS online documentation
E-Education Institute - Slope, Aspect, and Hillshade