PROBABILITY OF PRECIPITATION ESTIMATION USING SEVIRI DATA AND ARTIFICIAL NEURAL NETWORKS

The estimation of precipitation from geostationary satellite data has always been pursued because of their high spatial and temporal resolution and full disc coverage. The SEVIRI upgraded set of VIS/NIR/IR channels provides new means to address this challenge. As an initial step in that direction this work proposes a SEVIRI-based statistical algorithm to estimate Probability of Precipitation (PoP) for the UK area for summer daytime periods. This work moves on from previous work, presented at EUMETSAT-2004, where the capability to distinguish rain from no-rain of SEVIRI-like channels was assessed by considering the MODIS sensor. A MODIS–based algorithm is still used here in order to provide secondary support for validation of the main SEVIRI algorithm. The UK radar network provides precipitation data for calibration and validation. The statistical approach is based on Artificial Neural Networks (ANNs). A large supervised data set for June, July and August 2004 has been built. Radar precipitation estimation data (with 5-minute temporal resolution) have been collected from the Met Office Nimrod archive. SEVIRI data (between 8.00 UTC and 18.00 UTC, with 15-minute temporal resolution) have been provided by the EUMETSAT archive facility. MODIS data (AQUA and TERRA overpasses between 10.30 UTC and 13.30 UTC: around 100 cases) have been retrieved from the DAAC/NASA online archive. All the data are mapped on to the Nimrod 5 km grid. The analysis has been carried out in two parts: 1) around 12.00 UTC both SEVIRI and MODIS PoP estimators are defined, validated and compared, 2) between 8.00 UTC and 18.00 UTC a SEVIRI PoP estimator is defined and validated. The Equitable Threat Score (ETS) parameter is used as the skill indicator, with a data set independent from the one used during the ANN training and testing phases. The sensitivity of the SEVIRI-based algorithm to the diurnal cycle and different months is also investigated and discussed. The main results are that: 1) SEVIRI and MODIS PoP estimators perform in a comparable way, 2) The group of channels: VIS0.6 (0.64 µm), VIS0.8 (0.81µm), NIR1.6 (1.64µm), IR3.9 (3.9µm), IR8.7 (8.7µm) and IR12.0 (12.0µm) perform a summer rain no-rain classification with a mean ETS around 46% (for cloudy and clear sky pixels and for a dry to wet ratio around 10), 3) The largest contribution to this performance comes from the combination of VIS0.8 and NIR1.6 channels.