A new methodology for rain identification from passive microwave data in the Tropics using neural networks

International audienceThe detection of rainfall remains a challenge for the monitoring of precipitation from space. A methodology is presented to identify rain events from spaceborne passive microwave data using neural networks. We focus on BRAIN, the algorithm that provides instantaneous quantitative precipitation estimates at the surface, based on the MADRAS radiometer onboard the Megha-Tropiques satellite. A version of BRAIN using data from the Tropical Rainfall Measuring Mission (TRMM) Microwave Imager (TMI) has been used to compare several multilayer perceptrons (MLP) trained on different combinations of TMI brightness temperatures with the conventional GSCAT-2 algorithm approach used for rainfall detection. These classifiers were compared at a global scale to reference values from the TRMM Precipitation Radar (PR). They were also compared to ground measurements using two 1° × 1° dense rain-gauge networks from different climatic zones in West Africa to assess the influence of rainfall types. At the global scale the MLPs provide better Probability of Detection than the GSCAT-2 decision tree but tend to have a higher False Alarm Rate. While no unique solution exists given the strong regional dependence of the classifiers' performances, the screen based on the 19, 21 and 85 GHz channels provides the best detection results at the instantaneous scales. As to accumulated rainfall, the screen that exhibits the lower bias relative to the PR makes use of the 37 and 85 GHz channels. The evaluation over West Africa using 10 years of TRMM overpasses shows that MLPs are in better agreement with both the PR and the gauges than GSCAT-2. The MLP trained on the 37 and 85 GHz channels increases the Probability of Detection by nearly 35% compared to the former screening over the two studied regions. Better results are obtained in the case of organized systems