Get access

An error model for instantaneous satellite rainfall estimates: evaluation of BRAIN-TMI over West Africa



Characterising the error associated with satellite rainfall estimates based on space-borne passive and active microwave measurements is a major issue for many applications, such as water budget studies or assessment of natural hazards caused by extreme rainfall events. We focus here on the error structure of the Bayesian Rain retrieval Algorithm Including Neural Network (BRAIN), the algorithm that provides instantaneous quantitative precipitation estimates at the surface based on the MADRAS radiometer on board the Megha-Tropiques satellite. A version of BRAIN using data from the Tropical Rainfall Measuring Mission (TRMM) Microwave Imager (TMI) has been compared to reference values derived either from TRMM Precipitation Radar (PR) or from a ground validation (GV) dataset. The ground-based measurements were provided by two densified rain-gauge networks in West Africa, using a geostatistical framework. The comparisons were carried out at the BRAIN retrieval scale for TMI (instantaneous and 12.5 km) and over a ten-year-long period. The primary contribution of this study is to provide some insight into the most significant error sources of satellite rainfall retrieval. This involves comparisons of rainfall detectability, distributions and spatial representativeness, as well as separation of systematic biases and random errors using Generalized Additive Models for Location, Scale and Shape. In spite of their different sampling properties, the three rain estimates were found to detect rainfall consistently. The most important BRAIN-TMI error is due to the rain/no-rain delimitation which causes about 20% of volume rainfall loss relative to PR and GV. BRAIN-TMI presents a narrow PDF relative to GV and catches the spatial structure of the most active part of rain fields. The conditional bias is significant (e.g. +2 mm h−1 for light-moderate rain rates, −2 mm h−1 for rain rates greater than 8 mm h−1) and the overall bias is within 10%. The PR shows a significant underestimation for high rain rates with respect to GV. The proposed framework could be applied to the evaluation of other passive microwave sensors (SSMI, AMSR-E or MADRAS) or rainfall satellite products. Copyright © 2012 Royal Meteorological Society

Get access to the full text of this article