An evaluation of big data analytics in feature selection for long-lead extreme floods forecasting

A type of extreme disastrous floods are associated with a sequence of prior heavy precipitation events occurring frequently from over several days to several weeks. Transitional methods for precipitation clusters prediction usually rely on the measurement and analyses of meteorological variables. However while a short-term prediction of certain location depends only on variables in near spatial and temporal neighborhood, predictions with long lead time must consider variables in a long time window and large spatial neighborhoods, this means an enormous amount of potentially influencing variables and only a subset of them strongly relate to prediction. Processing a deluge of variables and discovering strongly relevant features pose a significant challenge for big data analytics. Finding influencing variables calls for automated methods of strongly relevant feature selection, which is what online streaming feature selection provides. In particular, online streaming feature selection, which deals with the stream of features sequentially added while the total data observations are fixed, aims to select a subset of strongly relevant features from the original feature set. In this paper, we apply four state-of-the-art online streaming feature selection methods for building long-lead extreme floods forecasting models, which identify optimal size of strongly relevant meteorological variables and confine learning the prediction model on the relevant feature set instead of the original entire feature set. The prediction models are evaluated and compared systematically on the historical precipitation and associated meteorological data collected in the State of Iowa.