A Neural Network Approach for Nonlinear Causality Analysis of Aerosol Influence on Physical Parameters Required for Tropical Cyclone Genesis over the Tropical Southwest Pacific Ocean

Tropical cyclones (TC) over the Southwest Pacific Ocean (SWPO) cause major disasters in Eastern Australia (EA). In the preliminary portion of the research Granger causality used to statistically determine the causal relationship between monthly mean aerosol optical depth (AOD) and thermodynamic (sea surface temperature (SST)) and dynamic parameters (specific humidity (SH), and relative vorticity (RV)) required for TC genesis over the tropical SWPO region for 1985 – 2015. Model results support the hypothesis that AOD is indeed helpful to predict SWPO SST, SH, and RV, but the causal direction did not goes the other way around. While this preliminary research approaches to granger causality detection based upon linear time series assumptions, many interactions in climate science is non-linear. We develop an approach to nonlinear granger causality detection using artificial neural network model to derive the AOD influence on physical parameters required for TC genesis. We propose a multilayer perception (MLP) neural network with backpropagation technique where the input to the network is the past time lags of all series and the output is the future value of a single series. For the estimation of MLP, we used a hierarchical group lasso penalty that automatically detects both nonlinear Granger causality and the lag of each inferred interaction. Finally, we used our MLP method to explore causal interactions between AOD with physical parameters required for TC genesis. To justify the performance, we carry out a set of experiments on individual TC events over the study period to prove that our proposed model is promising.