CLOUD DETECTION BY DISCRIMINANT ANALYSIS GERB and AVHRR case studies U. Amato+, L. Cutillo*, V. Cuomoo, C. Seriox +Istituto per le Applicazioni del Calcolo ‘M. Picone’ CNR, Napoli, Italy *Dipartimento di Matematica e Applicazioni, Università di Napoli ‘Federico II’, Italy oIstituto di Metodologie di Analisi Ambientale CNR, Potenza, Italy xDipartimento di Ingegneria e Fisica Ambientale, Università della Basilicata, Potenza, Italy GIST-17 Meeting, London, February 5th 2003 Plans to use GERB/SEVIRI data •Case Study: Desertification processes in Southern Italy •Methodology: Energy Balance at the Surface •Tools to be developed: (Among Others) Cloud Clearing and Cloud detection CLOUD DETECTION Physical methods Physical methods mainly based on thresholds evaluated by Radiative Transfer models Criteria for cloud detection often based on couples of reflectance/radiances at different wavelengths Multispectral and hyperspectral sensors potentially increase accuracy of cloud detection, but pose new challenges to the algorithm development CLOUD DETECTION Statistical methods Discriminant Analysis methods Nonparametric estimate of the radiance/reflectance density functions Transform of the radiance/reflectance multispectral components into new components (e.g., Principal Component Analysis, PCA; Independent Component Analysis, ICA) Classification by a classical Bayes rule Multispectral images Cloud mask Training set Multispectral images DISCRIMINANT ANALYSIS Nonparametric density estimation Data transformation Cloud detection Case study: GERB GERB-like data, format ARCH 60-minutes snapshots Full-disk Spatial resolution: about 33% of the 3x3 SEVIRI grid (833x833 pixels, 3Km x 3Km at the sub-satellite point) SW radiance ( < 4 mm) LW radiance ( > 4 mm) Test Latitude Longitude Time Day Train [-45o,+60o] [-20o,+60o] 16:00 Jun 21st 2001 Test [-45o,+60o] [-20o,+60o] 16:00 Jun 21st 2001 Success percentage (Linear Discriminant Analysis) Clear Cloudy Total Sea - SW 82.2 95.8 92.7 Sea - LW 86.6 56.5 63.4 Land - SW 82.5 79.6 82.0 Land - LW 83.9 52.4 78.8 Test Latitude Longitude Time Day Train [-45o,+60o] [-20o,+60o] 16:00 Jun 21st 2001 Test [-30o,+55o] [0o,+25o] 12:00 Feb 8th 2001 Success percentage (Linear Discriminant Analysis) Clear Cloudy Total Sea - SW 76.4 97.1 88.6 Sea - LW 85.7 37.8 57.6 Land - SW 98.3 44.4 85.7 Land - LW 84.9 66.9 80.7 Case study: AVHRR AVHRR onboard of NOAA 14 Full-disk Spatial resolution: 8 Km x 8 Km at the sub-satellite point 5 channels: 0.63 mm, 0.91 mm, 3.74 mm, 10.8 mm, 11.5 mm Test Latitude Longitude Day Train [-45o,+60o] [-20o,+60o] Dec 21st 2001 Test [+30o,+55o] [0o,+25o] Jun 21st 2001 Success percentage (NonParametric Discriminant Analysis) Clear Cloudy Total Land - 0.63 mm 93.0 100 94.6 Land - 0.91 mm 67.9 99.4 75.3 Land – 3.74 mm 29.0 72.5 39.2 Land – 10.8 mm 67.8 100 75.4 Land – 11.5 mm 85.8 75.7 83.4 Test Latitude Longitude Day Train [-45o,+60o] [-20o,+60o] Jun 21st 2001 Test [+30o,+55o] [0o,+25o] Dec 21st 2001 Success percentage (Linear Discriminant Analysis) Clear Cloudy Total Land - 0.63 mm 97.0 97.6 97.2 Land - 0.91 mm 66.8 99.9 74.6 Land – 3.74 mm 35.2 0.3 27.0 Land – 10.8 mm 72.9 99.3 79.1 Land – 11.5 mm 72.5 99.8 78.9 Perspectives To make density functions of radiance/reflectance least depending on time and location To choose a proper transform of multispectral data aimed at picking essential information and eliminating redundancies To merge physical and statistical models into a mixed model able to share benefits of both