Cloud detection using SEVIRI IR channels

Transcription

1 Cloud detection using SEVIRI IR channels & Luis Gonzalez Sotelino Royal Meteorological Institute of Belgium GERB Science Team London September / 19

2 Overview GERB Science Team London September / 19

3 Scene identification only relying on visible channels SEVIRI data affected by sun glint over ocean Sun glint saturating 0.6 & 0.8 µm channels Degraded cloud mask within sun glint area Cloud mask unavailable at night time GERB Science Team London September / 19

4 GERB aim is to study climate GERB products must remain stable Limited use of uncontrolled ancillary data Independence to NWP data Implementation of an IR cloud detection scheme instead of using MPEF or NWCSAF GERB Science Team London September / 19

5 Physics Physics Assumptions Scheme Initialization SEVIRI IR 10.8, 8.7 & 12.0 µm channels are most sensitive to clearsky & clouds SEVIRI IR 3.9 µm channel is most sensitive to low water clouds Clouds are characterized by lower radiances (temperatures) than clearsky surfaces (warmer) except for snow & sea ice surfaces Aerosols are generally lowering IR radiances IR radiances are varying with viewing zenith angle, history (precipitation, cloud shadow) and state of atmosphere (profiles) Visible & IR cloud masks will have discrepancies due to different measurement sensitivities GERB Science Team London September / 19

6 Assumptions Physics Assumptions Scheme Initialization Considering time series of pixel based BTs Temporal window for time series set to 60 days Samples in time series can be grouped into 3 classes: 1. thick cold clouds (low BTs) 2. thin or low clouds (high BTs) 3. clearsky conditions (highest BTs) Tails of upper classes are overlapping No realtime ancillary data such as NWP fields Cannot be applied to snow & sea ice surfaces GERB Science Team London September / 19

7 Scheme Physics Assumptions Scheme Initialization Perform a modified k means clustering: 1. Initialize the µn and σn for the 3 clusters 2. If initialization fails goto step 1 with 2 clusters and so on Classify all 60 BTs according to their nearest cluster with d(t, µn, σn) 4. Update µn and σn 5. Repeat from step 3 until all µn do not significantly change ( µn < 0.01 K) Metric d(t, µn, σn) = (T µn) 2 /σ n 2 + ln σ2 n if values in each class follow pn(t) = N(µn, σn) Initialization driven by physics (climatology) GERB Science Team London September / 19

8 Scheme Final classification (of the most recent sample): Physics Assumptions Scheme Initialization CLOUDY p 2 (T) CLEAR UNCONTRASTED p 1 (T) p 0 (T) ε 1 ε 2 S S 1 2 T [K] x 1 r x 2 Equality of the probabilities of no good classification: ǫ1 + S1 = ǫ2 + S2 Equality of probabilities of uncontrasted and false classifications: ǫ1 + ǫ2 = S1 + S2 S1 = ǫ2 and S2 = ǫ1 GERB Science Team London September / 19

9 Initialization Physics Assumptions Scheme Initialization Assume that clearsky class is wide Cloudy classes evenly distributed over remaining T range is only needed for starting the clustering Single cluster case associated to clearsky is estimated from climatology: 10 years of 6 hourly ERA 40 surface skin temperatures Ts Compute δt = T s (59) T s (2) at pixel level (x, y) (x, y) is the median of 10 years seasonal δt(x, y) GERB Science Team London September / 19

10 Initialization Physics Assumptions Scheme Initialization MAM [K] at 00:00 UTC GERB Science Team London September / 19

11 3.9 µm March at 00:00 UTC GERB Science Team London September (single channel) 8.7 µm 11 / 19

12 10.8 µm March at 00:00 UTC GERB Science Team London September (single channel) 12 µm 12 / 19

13 (single channel) Reference is MPEF & NWCSAF common cloud mask (March 2007 at 00:00 UTC) Day Band Common [%] POD [%] fcs [%] fcl [%] POD [%] fcs [%] fcl [%] POD [%] fcs [%] fcl [%] POD [%] fcs [%] fcl [%] fcs = false clearsky, fcl = false cloudy GERB Science Team London September / 19

14 (single channel) Limitations: Cloud masks sensitive to specific clouds Cloud masks cannot be simply merged 10.8 µm gives most consistent results Low water clouds systematically missed ( 1K) GERB Science Team London September / 19

15 Suggestions Suggestions NWCSAF algorithm uses threshold on: T10.8 T3.9 for low water clouds T12 T3.9 for low water clouds (ocean) T10.8 T12 for thin cirrus and cloud edges T8.7 T10.8 for thin cirrus Note that 3.9 µm channel only used at night GERB Science Team London September / 19

16 Suggestions Low water clouds improved detection (night): Use 10.8 µm cloud mask for reference Use joint 3.9 & 10.8 µm 1D clustering results For pixels with discordant 3.9 & 10.8 µm cloud masks: 1. Compute 2D MLE (µn,σn) on joint common: clearsky class low contrast cloudy class 2. Classify most recent sample pair T according to nearest 2D cluster with d(t, µn,σn) Metric: d(t, µn,σn) = (T µn) t Σn 1 (T µn)+ln Σn if values in each class follow pn(t) = N(µn,Σn) GERB Science Team London September / 19

17 cm 10.8 = cs, cm 3.9 = CL cm 10.8 = cm 3.9 = CL cm 10.8 = cm 3.9 = cs Suggestions 255 T3.9 [K] T 10.8 [K] 2D MLE assigns most recent pixel to cloudy GERB Science Team London September / 19

18 (3.9 & 10.8 µm bands) Reference is MPEF & NWCSAF common cloud mask (March 2007 at 00:00 UTC) Day Band Common [%] POD [%] fcs [%] fcl [%] POD [%] fcs [%] fcl [%] fcs = false clearsky, fcl = false cloudy GERB Science Team London September / 19

19 improvement for low water clouds applicable to day time (3.9 µm)? with other channel combinations (NWCSAF like)? during day time Multidimensional k means clustering? ECMWF surface skin temperatures should be converted to TOA temperatures according to atmospheric path Use of asymmetrical distributions pn(t) instead of N(µn, σn) Length of time series varying according to pixel GERB Science Team London September / 19