Cloud screening and snow detection with MERIS Rene Preusker, Jürgen Fischer, Carsten Brockmann, Marco Zühlke, Uwe krämer, Anja Hünerbein
Prelude The following algorithm has been compiled in the frame of the ALBEDOMAP project (the generation of global 16 day spectral albedo maps; see presentation of J. Fischer). It is therefore limited to land surfaces!
Outline Objectives Algorithm Likelihood of cloudiness blue band screening cloud edge processing snow restoration analysis of short term variability analysis of long term variability Summary
Objective, or: Why did we need a new cloud detection? There is a standard cloud detection in MEGS. It is using two prominent features of clouds: 1. clouds are white 2. clouds are bright from satellite. This is by far not enough since: 1. it misses many thin clouds 2. it treats often snow as clouds (same for sun-glint) 3. it misses partly cloudy pixel In 16 days averages (with 1-6 overpaths) only one missed cloud destroys any kind of average
Example for insufficient MERIS L2 cloud detection RGB MERIS L2
Example for insufficient MERIS L2 cloud detection RGB MERIS L2
Algorithm: Likelihood of cloudiness Clouds are white Clouds are bright Clouds are higher than ground Direct utilization of the O 2 absorption band measured by MERIS at 760nm
Oxygen-A band differential absorption radiance ratio between an absorbing and a window channel depends on photon path length photon path length is mainly determined by air-mass above the cloud = cloud top pressure small ratio == no/low cloud ratio close to 1 == high (opaque) cloud
Sketch of Implementation Refl 1 Refl 2 Refl 3 Cls are bright Cls are white Cls are high Refl 4 Refl 5 Refl 6 Refl 9 Refl 10 Refl 13 O2A Srf.-Press Wvl 11 Cloudiness Probability Estimator (based on huge amount of RTM simulations, ANN ) Likelihood of clouds
Intermediate Although the cloud probability processor is working much better than the standard L2 processing: 1. many, especially thin/broken clouds slip through 2. snow had often a high cloud probability --> Need for some additional spectral filter to find partially cloudy pixel and a snow restoration
Blue band thresholding Physical Background: surfaces are very dark in the blue (412 nm) exceptions are: Clouds Snow ( applicable Sunglint (not very high aerosol loadings (difficult atmospheric correction --> unwanted) Implementation: Simple threshold of 0.2 in (Rayleigh corrected) ρ
Snow restoration Why wanted: Cloud probability and blue band threshold often falsely identify snow as cloud Physical Background: snow shows absorption in SWIR, that other surfaces don't show. Exceptions: water, but this is very dark in NIR few desert sites, but these are very dark in UV (!! climatology ) some huge tropical clouds This absorption is not very pronounced for the MERIS channels, but (thanks to MERIS's excellent radiometric resolution) sufficient!
Snow restoration example MERIS norm. diff. snow index: mndsi := (13-14)/(13+14) If (mndsi > 0.02) and not dark and not (sub)tropics then SNOW
Cloud edge processing Why wanted: Cloud shadow can not be adequately atmospherically corrected partially cloudy pixel are more often close to cloudy pixel removal of adjacency effects (brightening) due to clouds Implementation: 4x4 neighborhood around clouds is automatically excluded cloud shadow is geometrically calculated from MERIS ctp and sun geometry
Post processing Although we squeezed out the spectral potential of MERIS we still missed many, especially thin/broken clouds! --> Need to extend the feature vector by a temporal dimension. (This is of course only possible when producing some kind of L3)
Analysis of 16 day (short term) temporal variability in the blue band Physical background: Most land surfaces are very dark in UV. Exception: snow --> small cloud contaminations are first seen in the UV Implementation (empirical): If variability in 16 day bin exceeds 12.5% then each ρ > (16_day_mean + 2 stdv) is excluded If variability in 16 day bin exceeds 25% then each ρ > (16_day_mean + 1 stdv) is excluded
Analysis of the (long term) temporal variability in the blue band This test is not an test on ρ! Instead it finds L3 means/albedos which are still cloud contaminated. Why wanted: 16 day (short term) variability is in some cases not applicable, in particular when the weather is not changing during that time period. Low 16 day variability because of constant partly cloud cover false negative Implementation: A L3 bin is assumed to be cloud contaminated if the white sky albedo at 412nm exceeds the annual median by 50%. (Plus some extra logic to ( snow avoid false positive by
Summary 1. MERIS is NOT a god instrument to detect clouds (missing SWIR an TIR channels) even though all standard MERIS algorithms need a preceding cloud screening or cloud detection. (OLCI will benefit from SLSTR!!!) 2. We developed an highly improved cloud detection using all spectral features some temporal information almost no textual information 3. The investigation of MERIS L3 16 day albedo (which is very sensitive) indicates high quality of cloud detection (however a comparison with thermal sensors and ground ( different based measurements may look