Meta-heuristic Bayesian networks retrieval combined polarization corrected temperature and scattering index for precipitations

Abstract

This paper proposes Bayesian networks (BNs) that combine polarization corrected temperature (PCT) and scattering index (SI) methods to identify rainfall intensity. To learn BN network structures, meta-heuristic techniques including tabu search (TS), simulated annealing (SA) and genetic algorithm (GA) were empirically evaluated and compared for efficiency. The proposed models were applied to the Tanshui river basin in Taiwan. The meteorological data from the Special Sensor Microwave/Imager (SSM/I) of the National Oceanic and Atmospheric Administration (NOAA) comprises seven passive microwave brightness temperatures, and was used to detect rain rates. The data consisted of 71 typhoons affecting the watershed during 2000–2012. A preliminary analysis using simple meta-heuristic BNs identified the main attributes, namely the brightness temperatures of 19, 22, 37 and 85GHz for rainfall retrieval. Based on the preliminary analysis of a simple BN run, the advanced BNs combined with SI and PCT successfully demonstrated improved rain rate retrieval accuracy. To compare the proposed meta-heuristic BNs, the traditional SI method, the SI-based support vector regression model (SI-SVR), and artificial neural network (ANN) were used as benchmarks. The results showed that (1) meta-heuristic BN techniques can be used to identify the vital attributes of the rainfall retrieval problem and their causal relationships and (2) according to a comparison of BNs combined with PCT and SI and artificial intelligence (AI)-based models (SI-SVR and ANN), in heavy, torrential, and pouring rainfall, models of BNs combined with PCT and SI provide a superior retrieval performance than that of AI-based models. Therefore, this study confirms that meta-heuristic BNs combined with PCT and SI is an efficient tool for addressing rainfall retrieval problems.