1. WP 2000 Improved Identification of Clouds Jane Hurley, Anu Dudhia, Don Grainger University of Oxford

2. Current Cloud Detection Colour Index (CI) Method is now used to detect cloud … A couple of caveats … however … microwindows have not been optimized – would be useful if CI ~ EF fails to detect cloud with cloud fraction < 30%

3. Cloud Identification: SVD Analysis of Clouds

4. Objective: To create and analyze RFM-simulated cloudy spectrum of varying cloud effective fraction EF to formulate a new cloud detection method using Singular Vector Decomposition SVD. Singular Vector Decomposition SVD is statistical technique used for finding patterns in high dimensional data; transforms a number of potentially correlated variables into a smaller number of uncorrelated variables (SINGULAR VECTORS) first SV captures the most variance … and each successive SV captures increasingly less variance Idea is to find singular vectors that describe clear and cloudy atmospheres and use them in cloud detection

5.  Use 2 nd half of A band because more sensitive to cloud presence RFM-simulated spectrum with EF = 0 and 9.0km tangent height and the corresponding first 8 Clear Singular Vectors SV clear Clear Singular Vectors

6. Need only first few SVs to well represent signal: First 3 SVs capture ~90% of total variance

7. Cloudy Singular Vectors Subtract off mean spectral radiance from Original signal; Use SV clear to do a Least Squares Fit (LSF) on Original-Mean signal; Subtract LSF from Original signal to get Cloud-Only signal.

8. Compare with Aerosol signature with same EF in the FOV: SVD-calculated Cloud-only signal Aerosol signal Do SVD on Cloud-only signal to get Cloudy Singular Vectors SV cloud

9. Use SV clear and SV cloud for given tangent height to do a LSF to mean-subtracted Original signal 15 km 12 km 9 km 6 km EF ≠ 0 → Non-zero fit coefficient to SV cloud ! Application to MIPAS data

10. Cloud Detection Method 1: χ 2 Ratio Use SV clear and SV cloud to do LSF of arbitrary spectrum. Use χ 2 error to measure goodness of fit.

11. Method 2: Ratio of Integrated Reconstructed Radiances Use SV clear and SV cloud to do LSF of arbitrary spectrum. Reconstruct cloudy and total radiance using LSF.

12. Comparison of Methods 4 Methods of Cloud Detection: Radiance Thresholding Colour Index SVD χ 2 Ratio SVD Ratio of Integrated Reconstructed Radiances Methods applied to RFM Data: Percent that method gets prognosis right Methods applied to MIPAS 2003 data: Percent agreement between methods

13. Future Work Finish selecting optimal microwindows for use with Colour Index Method – those that best correlate with EF. Finalize choice of thresholds for SVD Methods. Compare methods of cloud detection on large MIPAS dataset against known databases (ISCCP) etc to see what difference this makes. Do SVD analysis of ice clouds and implement this into an identification scheme. Hopefully will then have a cloud type identification scheme.

15. For RFM-simulated spectra …

16. For MIPAS data, not so sharp a distribution, obviously … but clearly a bimodal distribution  Can fit a Gaussian to ‘clear’ peak and set a threshold: thr = peak + 3st.dev.  Should pick up 99.5% of cloud, if truly Gaussian

17. Cloud Detection Method 1: χ 2 Ratio Use SV clear and SV cloud to do LSF of arbitrary spectrum. Use χ 2 error to measure goodness of fit. Fit given spectrum with SV clear → χ 2 clear Fit given spectrum with SV clear and SV cloud → χ 2 clear+cloud Consider ratio of χ 2 clear / χ 2 clear+cloud : χ 2 clear / χ 2 clear+cloud > 1 for cloudy spectra χ 2 clear / χ 2 clear+cloud ≈ 1 for clear spectra

18. Application to RFM-simulated spectra …

19. Method 2: Ratio of Integrated Reconstructed Radiances Use SV clear and SV cloud to do LSF of arbitrary spectrum. Radiance of cloud L SVcloud = mean(Σ i (fit coeff) i SV cloud i ), where i ranges over the cloudy SVs only and average over spectral points.  L SVcloud = 0 for clear spectra  L SVcloud > 0 for cloudy spectra Total radiance L SVall = mean(Σ i (fit coeff) i SV i ), where i ranges over all SVs and average over spectral points. Consider ratio L SVcloud / L SVall :  L SVcloud / L SVall = 0 for clear spectra  L SVcloud / L SVall > 0 for cloudy spectra